[CatIIIIIIII]

What and why

Authors need to ask about a paper without starting an editing task. Add a separate Paper Q&A pane below the editing composer and task conversation, with a resizable divider and the existing agent/model/effort picker.

• Start conversations with the plus button and return to earlier sessions through the clock icon menu. Each history entry has a delete action with confirmation, failure retry and draft cleanup. Running answers cannot be deleted; removing the active conversation selects another or creates a fresh empty session. Questions and answers are stored in SQLite, scoped to paper, viewer and session; local drafts and model/effort choices survive switching sessions and reloads.
• Give each answer a bounded snapshot of the current manuscript and recent completed turns. Reuse the configured Codex, Claude or GLM runtime; allow web lookup and preserve the chosen reasoning effort on retries.
• Deduplicate requests, keep failed questions retryable, recover interrupted questions after restart, and hold project operations while an answer is running.

Permission behavior: Q&A displays a Full access indicator and runs directly in the selected paper directory. Codex uses danger-full-access with approvals disabled; Claude/GLM use the full-access adapter path, including commands, file tools, configured settings and network access. Requested edits therefore affect the current manuscript directly and do not go through an editing task's review worktree. OS-level permissions still apply. Read-only title/review helpers retain their existing restrictions.

History is managed by Papilot and passed as recent-turn context to ephemeral agent calls; it is not a terminal Claude session.

This PR is independently based on main (0.8.9). It does not require #36 (note editing) or #37 (compact slider). Keep conversation following synchronized before scroll events from the shorter pane can disable it. Existing browser fixtures now answer the new Q&A endpoints and target the upper composer explicitly where necessary.

How it was checked

• npm run typecheck
• node node_modules/vite/bin/vite.js build
• 29 tests: node --import tsx --test tests/agents.test.ts tests/agent-picker.test.ts tests/paper-question-agent.test.ts tests/paper-question-routes.test.ts tests/paper-question-sessions.test.ts tests/paper-questions.test.ts
• node scripts/check-paper-questions.mjs — independent composition, model/effort selection, history, session switching, retry, draft persistence, divider input, deletion confirmation/cancel/retry, session fallback, draft cleanup, desktop/mobile.
• node scripts/check-ui.mjs — including Git menus, task lifecycle, responsive layouts and long-conversation following.
• node scripts/check-note-selection.mjs
• node scripts/check-task-model.mjs
• node scripts/check-appearance.mjs
• node scripts/check-attachments.mjs
• Full npm test and aggregate npm run test:ui were not run; the checks above were run individually.

Screenshots use a synthetic PDF and mocked answers. The image files are hosted on a separate branch; generated builds and DMGs are excluded from this PR.