obto-feedback

OBTO reviewer feedback — read it before you build

Reviewers mark up the live preview of an app inside DevZone (Mark up → draw a box, arrow, pen stroke or label → note → Save). Each item is a record in ai_app_feedback in the app's own domain with:

• note — what the reviewer wrote;
• annotatedUrl — a PUBLIC PNG of the page with the reviewer's marks burned in (attach it to a vision-capable model to see exactly what they mean);
• snapshotUrl — the clean page capture, viewport (desktop 1440×900 or mobile 390×844, full page), capturedUrl (which page);
• overlay.shapes — the marks as vectors in the snapshot's pixel space;
• status — open → proposed → in_progress → fixed | closed | dismissed, with history and a linked missionId when a fix ran as a mission.

How feedback reaches you (three channels, no one has to tell you)

1. The tools. obto_list_feedback({appName, domain?, status?}) lists open items (status:"all" for everything); obto_get_feedback({feedbackId, domain?}) returns one item with the overlay and history. Both are read-only and are scoped to your connection's tenant (super-users may pass domain).
2. The nudge. When an app has open feedback, the results of obto_read_app_map, obto_fetch_app_graph and search end with [obto-feedback] N open reviewer feedback item(s)…. Treat it as a prompt to call obto_list_feedback before changing anything.
3. The bridge. Every new item is posted to the Agent Bridge thread feedback-<app> (kind question, author = the reviewer). Agents that poll bridge_inbox_peek see it; bridge_thread_read({threadId:"feedback-<app>"}) reads the backlog. Inside DevZone the platform also injects an open-feedback digest into every agent turn.

The routine

• At the start of work on an app: call obto_list_feedback once. If items are open, say so in one line ("2 open reviewer items — want me to take fb_… first?"); do not nag, do not repeat it every turn.
• When the user's request touches an item, or asks what reviewers want: call obto_get_feedback FIRST, look at annotatedUrl, and locate the exact artifact/element the marks point at before editing (discover-before-fetch: obto_read_app_map / obto_fetch_app_graph / search, never a guessed id).
• Fixing: patch the app the normal way (obto-build-loop), verify with obto_capture_preview on the page the item names, and compare against the annotated screenshot. Then tell the user the item can be marked fixed from the DevZone Feedback drawer — the MCP tools are deliberately read-only, so status changes stay with the humans. In DevZone itself, "Fix it" stages a plan turn first; a run by the reviewer starts the mission that flips it to fixed.
• Closed / dismissed items are hidden from the default list; ask before reopening them. Never delete feedback records.