verify

Feedback Studio — verify a processed round

The tool's whole promise is that the agent edited the right element. This skill extends that promise past the edit: for each resolved comment, confirm the requested change is actually present in the served page — and reopen any comment whose change can't be confirmed, rather than leaving a false "resolved".

The one rule that prevents false verdicts

Derive what to check from the comment text + thread (the requested END state) — not from whether the old anchor still resolves. A correct fix often invalidates its own anchor (the typo'd snippet no longer exists; a deleted element resolves to nothing). An anchor that fails to re-resolve is only suspicious when the comment did NOT ask for that element's text/existence to change.

Steps

1. Read .feedback/comments.json (or the dir passed as argument). Collect the resolved comments — they are the verification targets. Note each one's page, anchor (snippet = the OLD state), text/thread (the asked-for NEW state), and shot if present (.feedback/shots/<id>.png — the pin-time BEFORE image; view it to ground what the element looked like when the reviewer pinned it).
2. Markdown comments (sourceFile set): verify directly in the source — read the file, confirm the requested edit is present at the right spot (grep -n the new wording; confirm the old wording is gone). No browser needed.
3. Web comments: ensure the page is being served (start the server in the background as the feedback skill does — same --dir/--proxy source the round used — if it isn't already). Then, per page:
   • With browser tools (load the chrome MCP tools via ToolSearch in ONE call; skip this path entirely if they're unavailable): open the page, and for each comment check the new state — page text via get_page_text/javascript_tool for wording changes; a screenshot judged against the ask for visual changes (scroll the element into view first). window.__kbfSelfTest() is a useful secondary signal for elements whose anchors should have survived (a styling-only improve).
   • Without browser tools: fetch the served HTML (curl) and check the new wording is present and the old is gone. Purely visual asks ("make this stand out more") can't be machine-verified this way — mark them needs human eyes, don't guess a verdict.
4. Record every verdict on the comment itself so the file stays the audit trail. With the server running:
   • Confirmed: POST /__feedback/api/comments/<id>/reply with {"author":"agent","authorName":"verifier","text":"Verified: <what was checked and found>"}.
   • Not confirmed (change absent, or the WRONG element was changed): reply with what you found instead, then PATCH /__feedback/api/comments/<id> with {"status":"open"} — the pin flips back live and the comment re-enters the queue.
   • Unverifiable by machine: reply Needs human eyes: <why> and leave it resolved. (No server running: make the same edits in the JSON file directly.)
5. Summarise for the user: ✅ verified / ❌ reopened (with what was found) / 👀 needs human eyes — grouped by page, with counts. Be faithful: a reopened comment is the skill doing its job, not an embarrassment to soften.