Auto-Verify Loop

When the visual editor dispatches a card of style/text edits to a coding agent (Tier-2 — the edits a page's data-pb-loc can't resolve deterministically enough for an in-place source patch), PortBay doesn't just trust the agent's "Done" status. Once the card reaches Done, the workbench:

1. Reloads the live preview (without replaying the workbench's own prototype overlay).
2. Asks the in-page agent to check each requested edit's computed value against the reloaded page (verifyEdits).
3. Counts any new console errors that appeared since the reload.
4. Appends a ## Preview verification section — PASS/FAIL, per-property expected vs. actual, the console-error delta — back onto the card file.
5. Drops the edits that verified from the workbench's pending ledger (only the ones that didn't land stay pending).

This is what the Verify report modal (opened from the "View report" link in the footer while a card is watched) shows: a PASS/FAIL header, a per-property table, and — when a screenshot backend is available (macOS) — before/after thumbnails of each edited element.

Persisted history

Every completed run — a dispatched card reaching Done, or a manual re-run (below) — is written to .portbay/verifications/<id>/ in the project: the requested-edit spec, the per-property results, the console-error delta, and copies of the before/after element screenshots. Like visual edit history, this is plain JSON on disk (manifest.json + a shots/ subdir), not a hidden database — it version-controls with the project if you choose to commit .portbay/, and survives a PortBay restart.

The workbench's toolbar has a Verifications button (shield icon, next to History) that opens the run list: pass/fail, edits checked, new console errors, and which card drove the run, newest first. Runs are kept for 30 days (newest 30 kept), the same prune discipline as visual edit history.

Clicking a run opens the same Verify report view used for a live watch, over the stored data.

Re-running a stored verification

A stored run's requested-edit spec (the selectors/properties/values it checked) is enough to re-check it against whatever the page looks like right now — no card, no dispatch, no reload. Use Re-run on a row in the Verifications panel, or the Re-run button in the detail view. The result is persisted as a new run referencing the original (so the history stays append-only), with kind: "rerun" distinguishing it from a card-driven auto-verify. A re-run has no "before" reload to diff console errors against, so its new-error count is always 0.

Reading run summaries over MCP

The portbay_verify_runs MCP tool returns read-only run summaries for a project — id, timestamp, pass/fail, edits checked, new console errors, and the card path, without the per-property rows, spec, or screenshot paths (open the app's Verifications panel for the full detail). It's part of the opt-in verify toolset, alongside browser:

portbay-mcp --toolsets verify
# or combined with the browser companion
portbay-mcp --toolsets tasks,verify

Like the browser companion tools, portbay_verify_runs only exists on a Pro build compiled with both the tasks and visual-editor features — it needs the task board (a card is what triggers a run) and the visual editor (the verify loop itself).