Send visual context to an agent

Pinar writes plain Markdown and HTML to the clipboard. The text includes readable annotations plus a fenced pinar-visual-context JSON block. Paste both as one unit; the structured block is the machine-readable source of truth.

If the bundle lists an absolute Screenshot path, the local agent should open that single image; numbered badges are overlays. Warnings such as screenshot_missing, helper_unavailable, or viewer_unavailable describe degraded delivery but do not invalidate the comments and DOM context.

The Chrome extension never types into the agent composer. After Command/Ctrl/Alt+Enter, paste the clipboard yourself into Cursor, Claude, Codex, or Grok. The text begins by saying the pin notes may ask for a change or an explanation, and to treat selector and DOM path as complementary locators, followed by a fenced pinar-visual-context JSON block. If a Viewer URL is included, fetch it only when those details are not enough.

Treat captureId and pinId as identity, not labels to rewrite. Visual Context currently encodes schemaVersion 1; parseVisualCapture rejects a missing captureId and any schemaVersion other than 1 or the legacy 0. Follow only what the pins describe. If the person never pasted, ask them to copy again from Pinar instead of reconstructing pins from memory.

• Paste the whole clipboard into the agent; do not retype comments or invent a new captureId.
• Confirm the pasted text still contains a closed pinar-visual-context fence before you start editing code.
• If nothing was pasted, ask for Command/Ctrl/Alt+Enter in Pinar and follow only the pin notes.