Skip to content

SESSION_STATUS.md

Latest commit

 

History

History
83 lines (60 loc) · 9 KB

SESSION_STATUS.md

File metadata and controls

83 lines (60 loc) · 9 KB
Raw

Study Saga — Session Status (as of 2026-08-09)

Everything below is halted (laptop overheating). This is a handoff doc so work can resume cleanly. Nothing is currently running — no dev servers, no background scripts.

1. Cloudflare Deployment — ✅ DONE, live

  • Live at https://study-saga.pages.dev
  • Flask backend ported to Cloudflare Pages Functions (cf-pages/functions/), sessions moved from in-memory dict to Cloudflare KV (GAME_SESSIONS namespace).
  • Hint-credit system shipped: 3 Simple hints + 1 Deep hint per game, credits earned from correct answers (+1) and victories (+2) unlock more of either.
  • Per-level correct/incorrect results screen shipped (victory/defeat screens).
  • Fixed a real data bug: "7 - (-2)" had the wrong isCorrect flag in data.json (fixed in both backend/data.json and cf-pages/functions/_lib/data.json).
  • GitHub CLI (gh) and Cloudflare Wrangler CLI both installed and authenticated this session.

2. Git / Repo — ✅ DONE

  • Repo moved to https://github.com/TovTechOrg/study-saga (main branch), now the project's main repo.
  • Old repo (8GSean/study-saga) still exists but is no longer primary.
  • Commit authorship on your 4 commits corrected to Talia Kohen <Talia.Kohen@gmail.com> (the 4 older commits by Avihay Baratz and Yonatan Perlin were left untouched — not yours to reattribute).
  • Local git remotes: origin = 8GSean/study-saga, tovtech = TovTechOrg/study-saga.
  • The UI/UX work below (section 3) is NOT committed or pushed yet.

3. UI/UX Overhaul — 🔨 IN PROGRESS, code done, verified working, not yet redeployed

Fixing 7 GitHub issues Sean filed against the live deploy (TovTechOrg/study-saga issues #1-7): viewport overflow, inconsistent button styling, oversized typography, quiz layout/math rendering/accessibility, low-contrast character art, missing navigation. Full audit + plan saved at C:\Users\talia\.claude\plans\structured-stargazing-whisper.md.

All code changes are done:

  • cf-pages/public/static/css/style-neural.css — fully rewritten: responsive breakpoints, clamp() typography, focus-visible states, .answer-btn/checkbox styling, sprite glow, .game-nav styles, dead duplicate CSS block removed.
  • cf-pages/public/index.html — nav bar markup, ARIA (role="dialog", aria-modal, aria-labelledby) on both modals, KaTeX CDN tags, dead script.js reference removed, cache-bust bumped to ?v=5.
  • cf-pages/public/static/js/game-simple.js — duplicate realm-name bug fixed (eyebrow now says "Realm" instead of repeating the name), keyboard-accessible syllabus cards (tabindex, Enter/Space), focus trap + closeQuizModal()/deactivateModal() wired into both modals, #close-quiz-btn/Escape working, math normalizer + KaTeX render wired into question/answer text, nav bar wired (shows on syllabus-select/combat, hides on main menu, shows realm name in combat).
  • Deleted confirmed-dead files: cf-pages/public/static/css/style.css, cf-pages/public/static/js/script.js.

A real bug was found and fixed during verification — worth knowing if anything similar breaks later: the focus trap's initial .focus() call was being silently ignored by the browser. Root cause: getComputedStyle() on the quiz modal still reported visibility: hidden even a full animation frame after adding the .active class, because the modal's opacity/visibility change goes through a CSS transition (transition: opacity 0.4s ease, visibility 0.4s ease) that hadn't resolved yet — calling .focus() while the browser still considers the element non-focusable is a no-op, not an error. Fixed by waiting for the actual transitionend event (with a 450ms setTimeout fallback in case it never fires) instead of guessing at frame timing.

Verified working via a Playwright test script (backend/ui_fixes_verify.py, kept in the repo, uses backend/.venv which already has Playwright installed):

  • ✅ No horizontal scroll at mobile (390×844), tablet (768×1024), or desktop (1920×1080) on main-menu, syllabus-select, or combat screens.
  • ✅ Nav bar visible on syllabus-select/combat, shows realm name (e.g. "BIOLOGY") in combat.
  • ✅ Answer buttons are properly styled (neon dark theme, confirmed via computed background-color, not native gray).
  • ✅ Focus trap holds correctly — 12+ Tab presses cycle only within the modal, wrapping at both boundaries.
  • ✅ Escape closes the quiz modal correctly.
  • ⚠️ Realm-name-appears-twice check reported a false positive — that was a flaw in the test script (naive substring counting; "Biology" legitimately appears in both the <h3> and the "Biology realm" description text), not a real bug. The actual fix (eyebrow no longer repeating the name) is confirmed working by inspection.
  • ❓ Math/KaTeX rendering not actually confirmed yet — the test script's "hunt for a math question" loop has a bug: after pressing Escape to skip a non-math question and reopen with a new #attack-btn click, the quiz modal doesn't advance to a new question, causing later page.click("#attack-btn") calls to time out (the modal's already open, click hits the modal backdrop instead). This is a test-script issue, not necessarily a game bug — but it means KaTeX rendering is still unverified, and should be checked before considering this fully done.

Not started:

  • Fix the math-rendering test loop (or just manually check a math question in a browser) to confirm KaTeX actually renders.
  • Redeploy via wrangler pages deploy public --project-name study-saga --branch main.
  • Commit + push to TovTechOrg/study-saga.

4. Firebase Accounts/Gacha Feature — ⏸ PAUSED, blocked

Plan for this feature (auth, cross-device saves, points economy, gacha upgrades, potions/attack-power) is no longer saved anywhere as a plan file (the plan-mode file was overwritten by the UI/UX plan) — would need to be re-derived from this conversation's history if picked back up, or re-planned fresh.

  • Firebase CLI installed, logged in as taliakohen@tovtech.org.
  • Blocked on: firebase projects:create kept failing on "Callers must accept Terms of Service" — root cause turned out to be Google Cloud requiring 2-Step Verification (2SV/MFA) on the account, not actually a ToS issue (misleading error).
  • You clicked "Enable MFA"/"Turn on 2SV" but we never confirmed whether it completed.
  • Next step when resuming: confirm 2SV is active, then retry firebase projects:create study-saga-game --display-name "Study Saga".

5. LLM Bakeoff / Generator-Judge Architecture — 🔨 IN PROGRESS

Decided, strongly confirmed: Gemini as content-generation model. At 94.3% coverage of the BCP (biology/chemistry/physics) full bakeoff:

Subject Avg score n Below 7/10 floor
Biology 9.56 200/200 (complete) 3.0%
Chemistry 8.87 200/200 (complete) 14.5% (the one open thread)
Physics 9.52 166/200 4.8%
Overall 9.31 566/600 7.6%
  • Math syllabus bakeoff: 200/200, complete, separately.
  • BCP bakeoff progress since last full check: grew from 431→492→566/600 across several unattended runs (auto-resumes through Gemini quota walls). Last known state: not currently running; 34 physics questions remain. Re-run backend/bcp_full_bakeoff.py to continue whenever convenient — it's resumable and picks up where it left off.
  • Physics turned out to NOT be the weak subject (an earlier prediction that didn't hold up) — it's now essentially tied with biology for best-performing. Chemistry's 14.5% below-floor rate remains the one real open question — consistently worse than the other two subjects across the whole sample, unexplained.

Still open: Groq vs. Gemma as judge — no real verdict yet. Confirmed the main bakeoff data structurally can't answer this (hint_judge.py's fallback function only ever calls Gemma when Groq fails, never both). Reliability-wise, the split has evened out to ~50/50 (152 Groq / 155 Gemma of 307 tracked calls) — Groq's earlier rate-limiting looks transient, not fundamental. But reliability ≠ quality. Two scripts exist to actually answer the quality question, both still unfinished:

  • backend/finish_groq_vs_gemini_judged.py — scores the existing unscored 100-question Groq-vs-Gemini generator comparison. 7/100 done, resumable.
  • backend/groq_vs_gemma_judge_headtohead.py — the actual Groq-vs-Gemma judge quality test (30 sampled hints, judged independently by both). Never run at all.

Both need GROQ_API_KEY/GEMINI_API_KEY (already in backend/.env) and may hit the same Gemini quota wall as the main bakeoff.

Quick resume checklist

  1. UI/UX (closest to done): manually verify KaTeX math rendering in a browser (or fix the test script's loop bug), then wrangler pages deploy public --project-name study-saga --branch main, then commit + push to TovTechOrg/study-saga.
  2. Re-run bcp_full_bakeoff.py to finish the last 34 physics questions, then run the two judge-comparison scripts to finally settle Groq vs. Gemma.
  3. Confirm 2SV status on taliakohen@tovtech.org, resume Firebase project creation for the accounts feature.