[cu1ch3n]

What and why

Authors can request an AI review through the existing skill composer and read concise suggestions beside the relevant PDF passages. Each review builds and preserves a PDF from its task worktree; manuscript text stays unchanged.

• Keep findings in the review task until the author bookmarks individual annotations onto the manuscript. Copies are reanchored against the current PDF, deduplicated on retry, and retained when the original finding or task is removed.
• Keep personal notes private until explicitly quoted into an editing task. Reuse one compact editor for creating notes, editing PDF popovers, and the note library, with direct text editing and save/close through Enter, Escape, or clicking outside.
• Add left/right keyboard navigation between AI suggestions, a neutral bookmark action, and explicit task archiving.
• Ask the reviewer for one short, actionable sentence per finding, usually anchored to 1–3 lines, plus a one-sentence summary. Update the user guide to match the workflow.

Reviews use selectable PDF text, including extracted equations and captions. Annotations are stored in Papilot and displayed over the PDF; downloaded PDFs do not embed them.

How it was checked

• npm run typecheck
• npm test (CI on Node 22 and Node 24).
• All checks in npm run test:ui passed. The skill-library expectation was updated for the eighth built-in skill, then that check and the remaining attachment/account checks were rerun successfully.
• Desktop and mobile annotation screenshots inspected locally. UI regressions cover unsaved-note navigation and concurrent background refresh during suggestion restore.
• Backend coverage includes task PDF isolation, annotation ownership/CSRF, selective copying and retries, reanchoring after manuscript changes, retaining copied notes, and recovering preserved review PDFs after a repository move.
• Review engine source is unchanged; its standalone test suite was not rerun.

Integrated the latest main, preserving its skill icon metadata and settings changes.