# Study Desk V2 — progress and resume notes Updated 2026-11-12. Branch: **v2**. Original V1: **main**, preserved at `7bd0a9a`. The owner authorized publishing progress and `v2` now tracks Forgejo `origin/v2`. No production deployment has been performed. ## Product and architecture This is a rebuilt app using React + Vite + TypeScript, Hono/Node HTTP server, built-in `node:sqlite`, SQLite FTS5, local document storage and a persistent document queue. One process, one data volume, one administrator. Anonymous visitors get public read-only shares and browser-local study progress. Administrator progress stays server-side. The previous agent implemented the core product and committed it through `c549c34`: classes/folders, CRUD/duplicate/move/search, quiz sessions/results/history, FSRS flashcards, imports/exports/templates, documents/extraction/retrieval/OCR, provider adapters, editable AI drafts, tool-based assistant actions, sharing, analytics and the responsive themed UI. User feedback fixes are already included: larger clickable flashcards, arrow keys, streak effects, persistent review sidebar, live quiz score, shared quiz timers, SATA credit details, type-grouped materials, copyable import instructions and assistant library links. ## Current continuation - Confirmed original 162 tests pass; sandbox networking initially blocked the mock provider. - Added encryption-key corruption/missing-key refusal, static-path boundary validation, private embedded-media authorization and share-revocation checks. - Full backups include the generated encryption key and fail on unreadable files. - Added backup restore, media authorization, key durability, static-path traversal, V1 import and failed-migration rollback tests. The new V1 test checks source bytes are unchanged and state is preserved. - Added Dockerfile, Compose, .dockerignore, README and deployment/migration/AI documentation. - Updated obsolete ChatGPT subscription wording after checking official documentation: account integration exists for eligible clients, but hosted app approval for this private Unraid web deployment has not been established. API integration remains the working path. - Terra completed isolated UI/workflow and visual QA; see results below. - Guarded quiz exit/submission when the latest save failed, and prevented repeated automatic timer submissions after a failed request. - Preserved the V1 card elapsed-days scheduling field during import. ## 2026-11-12 additions - `.forgejo/workflows/build-v2.yml`: pushes to `v2` run typecheck + tests inside the Node 24 image (new Dockerfile `test` stage), build and smoke-test the production image, then publish `study:v2` and `study:v2-`. Never `:latest`. Verified locally with Docker Desktop; the workflow itself has not run on the Forgejo runner yet (nothing pushed). - Drag-to-reorder (`web/src/ui/Sortable.tsx`, pointer events so touch works, keyboard with arrows/Home/End and live announcements) for class materials, folders, questions and cards. New endpoint `POST /api/classes/:id/items/reorder`. - Timed exams use a wall-clock `deadlineAt` stored in attempt settings. The server refuses late answers (409 `time_up`), submits overdue exams lazily and before every authenticated GET, and caps time used at the limit. Shared exams use a local wall-clock deadline too. - Shared quiz sessions freeze their questions (with option order) in browser storage, with a fingerprint to tell the visitor when the owner has since edited the quiz. - Scanned PDFs are read in 8-page batches split with pdf-lib, cached per batch, with real page numbers. OpenAI-compatible adapter now sends PDFs as Chat Completions `file` parts when the model is marked as PDF-capable. - Semantic search: migration `2_embeddings` (`chunk_vectors`), embedding support in the OpenAI/Gemini/Ollama/compatible adapters, `server/services/semantic.ts` (background indexing, in-memory cosine search, reciprocal-rank fusion with FTS), Settings → AI section. - Browser QA on an isolated `.test-data/qa` instance (port 3742, `study-qa` launch config): drag and keyboard reorder persisted; 375px class page has no overflow; semantic search enabled from Settings and indexed; a 1-minute exam counted down on the wall clock and auto-submitted at 60 s; a shared session kept original questions after the owner edited them. ## 2026-11-15: first revision (owner's 21-item list) Implemented and verified on an isolated instance (`.test-data/rev1`, port 3743): DeepSeek provider (`server/ai/providers/deepseek.ts`); 7 reasoning-effort levels mapped per adapter; the "does not support tool calling" errors were DeepSeek rejecting forced tool choice in thinking mode: compatible servers now get the call requested in the prompt, and only a genuine "no tools" 400 reports unsupported. Cached-input prices (migration 3) and cache-write /off-peak aware costs. Built-in embeddings (`server/ai/localEmbed.ts`, bge-small, default when nothing is chosen). Assistant turns run server-side independent of the browser (re-attach via `GET /api/ai/chat/:id/stream`, stop via `POST /api/ai/chat/:id/stop`); drafts tray in the sidebar/Today. Quiz start screen, navigator ring/legend fix, timer only for timed exams, 7-point letter grades, numeric questions (migration 3 rebuilds `questions` for the new type; answers are compared numerically, units ignored, optional ± tolerance). Dialog footer padding (`safe-pb-3`), keyboard hints removed, small assistant button, full-width assistant page, model switcher and @-tagging. Editable generator instructions and built-in presets (builtin = 2 marks an edited built-in), separate chatbot import instructions. Copy buttons work over plain HTTP inside dialogs. Class/folder shares include processed documents. Live DeepSeek (`deepseek-flash` only) checks passed: discovery, quiz generation at default and Max effort (numeric items produced, answer keys correct with thinking on), multi-step tool turns, follow-up turns, tagged quiz to flashcards, an answer finishing after disconnect; zero errors in Usage. Docker image builds; ONNX Runtime and the embedding model run in the container. The image grew to about 1.3 GB because of ONNX Runtime. ## 2026-11-15: second revision (owner's 19-item list) Quiz generation: all quizzes NCLEX; "Practice exam" = 50 questions; calculation questions only when the source mentions calculations (max 5 in an exam); broad topics (prompted, plus `consolidateTopics` merging narrower labels); option order shuffled. The Python check uses at most 3 runs and ignores cosmetic remarks. Assistant: text, tool steps and cards stream in the order they happen (`MessagePart`), paragraphs are separated between steps, optional "Show thinking" box (persisted, fixed height, follows the newest text), "+" opens a material picker (no more flashing "@"), library items can be dragged into the panel, shorter placeholder. "View questions and answers" removed from the quiz start screen. Built-in embedding model baked into the image (`scripts/fetch-embedding-model.mjs`, `LOCAL_MODELS_DIR`), sharp stubbed when its native binary will not load, and the real load error shown in Settings. Python runner also blocks network and process modules in JavaScript (Node 24's permission model does not block network). New: MCP servers with per-server instructions and per-area switches (`server/ai/mcp.ts`, `runChatWithMcp` in `service.ts`); per-share AI switches (migration 4, `shares.ai`; `POST /api/public/:token/explain|chat`, limited per device via a signed visitor cookie plus a per-link cap; MCP servers join only with their "Share links" switch on); flashcard Standard mode with "repeat missed cards until known" next to Spaced repetition on the deck page; streak visuals restyled in the app palette (marker highlighter, accent bar, mono numerals); document Side by side / Original view with generated previews for Word and PowerPoint. Item 13 (pictures in PowerPoint files) was verified only: they are not read; see docs/AI.md. Tests: `tests/revision2.test.ts` (mock MCP server over Streamable HTTP and SSE, area switches, instructions in prompts, exam length, calculation allowance, topic merging, message part order, previews, share-link AI scope and rate limits). ## 2026-11-15: third revision (owner's 13-item list, plus exams and clarifying questions) Built-in Parallel web search (keyless, optional sign-in) with its own settings page; OAuth for every MCP server (`server/ai/mcpAuth.ts`); friendly tool steps (`web/src/features/ToolSteps.tsx`, `server/ai/activity.ts`) with ids stripped from results; thinking box 50% taller; per-chat tool switches; share-link assistant button matches the owner's and hides during a quiz or deck; share-link documents readable in the browser (original or text, page by page); live updates over SSE (`GET /api/events`, drafts and library); quiz results show a % ring coloured by band (rainbow at 100, green 93–99, blue 85–92, yellow 77–84, red below 77); Explain opens beside the quiz on wide screens and reads the cited source pages; OpenUI blocks in all chats; exams within classes (migration 5 adds `classes.group_label` and `conversations.settings`); assistant clarifying questions (`ask_user`, `web/src/features/AskCard.tsx`). Item 13 (beta container losing its data on update): the code keeps the database and secret key in `/data`, and compose.yaml mounts `${STUDY_DATA_DIR:-./data-v2}:/data`. Data is lost only when the container runs without that mount (the image's `VOLUME` then gives each new container a new anonymous volume). Map `/data` to a host path in the Unraid template. Tests: `tests/revision3.test.ts` (web search, OAuth with a mock authorization server, live events, cited pages, shared document text, genui prompt, clarifying questions, exam-scoped generation). ## 2026-11-15: fourth revision (owner's 12-item list) Interactive blocks encouraged by default (and never practice questions in Explain); Generate dialog's practice set / exam switch removed (the preset covers it); Explain beside the question on screens from 900px wide, moving the question left while open; breathing indicator and live stages on the assistant's tool steps, past-tense lines when done; no icon on Generate; the assistant asks before searching the web when the class materials do not cover a question; clarifying questions offer the class's exams and never ask about format (always NCLEX); Python check notes are green when every check passed, applied fixes neutral; share pages grouped by exam, then Quizzes, Flashcards, Notes, Class Materials (empty groups hidden, same order on the class page); "Documents" renamed "Class Materials"; OpenRouter provider (`server/ai/providers/openrouter.ts`) replacing the Ollama type (migration 6 turns Ollama connections into OpenAI-compatible ones at `/v1`). Tests: `tests/revision4.test.ts`; OpenRouter cases in `server/ai/providers/providers.test.ts`. Follow-up the same day: visitors of class and exam links can generate quizzes (default and maximum 50 questions, 25% select-all, optional 5 med math questions) and flashcards from chosen class materials (flashcards up to 75), 4 quizzes and 2 flashcard sets per link per day with a reset in the share dialog (`server/routes/shares.ts`, `web/src/features/SharedGenerate.tsx`, migration 7). Sets show in Drafts while written, then save into the class or exam automatically, and stay when the link is deleted. The owner's Generate dialog gained "Include med math" with a count (default 5). ## 2026-11-16: fifth revision (owner's 18-item list) Explanatory filler removed across the UI (share page, Generate dialogs, start screens, settings, flashcards); share-link "In progress" list for sets being written, finished sets join the normal lists with a "New" tag for two hours; visitors see "Verifying questions"; exam mode has no question-count slider (an exam is always the whole quiz); newest exam first on share pages; malformed JSON repaired by the model, retried, then the whole job restarted once; failed sets dismissable on the link only; plain-text copy everywhere; generated sets get their own short name and description; atomic flashcards without hints; a known / still learning counter in standard flashcard study; the assistant never answers from general knowledge (search or say it has no information); MCP and web search tools load automatically; share-link AI expires after a week, can be limited to the US or Arizona (`server/services/geo.ts`, DB-IP Lite) and is locked by abuse protection (migration 8). Follow-up the same day: the quiz check no longer hands over a quiz with known problems. It can rewrite a whole question, removed questions are replaced, a second independent review reads the whole quiz for badly written items (a writer changing its mind mid-stem), and the quiz is checked again until it passes (at most three rounds). Speed (timed live on a 50-question quiz): output limits now leave room for reasoning (calls at "max" were being cut off at 16k tokens), questions are planned before writing so batches can run four at a time without overlapping, replacements are planned and parallel too, and the two reviews run side by side. 50 questions: 7 min 21 s, 0 duplicates, all checks passed. Settings → AI now has separate Planning, Writing, Python check and Second review rows under Content generation (each inherits until set), and shared quizzes say what Practice and Exam mean under the mode toggle. Tests: `tests/revision5.test.ts`. Live check on the rev1 copy with DeepSeek flash: a visitor's 10-card set was written, named ("Heparin, Furosemide, and Dosage Calculations"), saved and listed once with the New tag; cards were single facts with no hints. The rev1 copy's 401 was its stored DeepSeek connection still holding the old key; it was updated from `.env`. ## Verification status Baseline: 162 tests passed, TypeScript passed and production web/server builds passed. Added six regression tests. Full `npm run check` passes: **281 tests** (generation steps, 2026-11-16; 280 after the check loop and second review; 274 after the fifth revision; 261 after share-link generation, 2026-11-15; 257 after the fourth revision; 247 after the third, 208 after the second, 181 on 2026-11-12), TypeScript and both production builds. The Docker `test` stage runs the same 181 tests on Node 24. Tests were run with loopback access for the mock provider. The earlier agent manually exercised quiz completion/results, flashcards, streak effects, mock AI generation through the assistant, editable drafts, and a 375px home page without horizontal overflow. Treat these as historical evidence; record Terra's fresh results below. Docker Desktop is now installed. Image build and a disposable non-root container passed login, document/card persistence across restart, full-backup ZIP, anonymous sharing and private API denial. Still unverified: Unraid filesystem permissions/proxy, paid-provider inference and real ChatGPT OAuth/eligibility. The installable app caches its shell, not private study data; full offline administrator study is not implemented. ## Remaining work 1. On Unraid later: execute the deployment/restore checklist in docs/DEPLOYMENT.md. 2. With owner-supplied provider connections: verify live discovery, streamed chat, generation, OCR and task routing. No real-provider calls should be inferred from mock test results. 3. Future optional work: approved ChatGPT account OAuth, more robust offline studying and (semantic embeddings were added on 2026-11-12, the built-in model on 2026-11-15). ## Resume safely 1. `git status --short`, read this file and recent commits. Do not overwrite user changes. 2. The ignored `.env` currently selects `.data/`, containing dev/demo content that the user has tried. Never print its secrets or reset/seed it casually. Use a new `.test-data/...` directory and a separate port for destructive QA. 3. `npm.cmd run check` (tests require localhost networking for their mock AI server). 4. `npm.cmd run dev` for Vite + API, or `npm.cmd run build` then `npm.cmd start` for the built app. See README and .env.example. All automated test databases are disposable copies. 5. Keep focused commits on `v2`. The owner has now authorized pushing progress to this branch; Git Credential Manager works. Never force push or publish/deploy to production implicitly. ## V1 comparison and follow-up usability work See `docs/V2-COMPARISON.md` for the requirement matrix, V1 differences and remaining priorities. Fixed the flashcard entrance/exit CSS conflict; restored touch grading and shuffle-remaining; reset review queues on scope changes; serialized note autosaves, protected navigation with unsaved edits and added an explicit retry action. Verified a note survives immediate navigation and reload, and shuffling keeps browse progress at 2/4 rather than restarting. The owner's rationale check passed using a synthetic three-question quiz with per-option explanations: correct, incorrect and partial SATA feedback all show explanations inline under their options. Results group explanations with each question. Screenshot is ignored at `.test-data/rationale-check/correct-inline.png`. The temporary test server is stopped after QA. ## Fresh UI QA findings Terra tested through Edge in an isolated volume on port 3731. Observed successful login, class creation, quiz/question editing, practice start, save/exit, resume, answer check and 100% results; deck/card creation, browse flip, FSRS review, pasted question import; desktop and emulated 400×933 mobile layouts in both themes. No obvious clipping or overlap observed on those screens. The test server was stopped and the temporary password removed. Visual finding: secondary labels appeared faint, particularly in dark mode. Adjusted the shared `ink-3` colors. Computed WCAG contrast against paper/surface/sunken backgrounds: light now **5.15 / 5.38 / 4.76** (previously 3.48 / 3.63 / 3.21); dark now **7.42 / 6.86 / 7.71**. These ratios cover these token pairs, not a full accessibility audit. Screenshots remain ignored under `.test-data/ui-terra/`: desktop-dark-quiz.png, desktop-light-quiz.png, mobile-dark-quiz.png, mobile-light-quiz.png and mobile-light-session.png. Post-fix Terra recheck: desktop light/dark screenshots show improved secondary-text legibility (`postfix-dark-desktop.png`, `postfix-light-desktop.png`). An Edge InPrivate visitor opened the isolated shared quiz without login, completed practice, saw its rationale and reached 100% results. Shared-deck UI, long-label layout stress tests and post-fix mobile contrast were not revisited in this follow-up. Public deck data isolation/revocation is covered by API tests. No known functional failures were found in the UI paths exercised. This is sampled QA, not an exhaustive audit of every layout and browser. Terra's isolated server is stopped.