Claude Project Tracker

activeShipped #147 and #148: official College Board score-report parsing, and vision extraction for images on gemma4. Closed as done: #150 (BRD23-B vision), #21 (real PDF parser, REQ-19.7 — deferred since June), #136 (web nav), #147 (Firebase deploy blocker, fixed days ago). Why the parser failed on a real PSAT: not a regex that needed loosening — we had modelled the wrong thing. CB reports contain NO per-skill correct/total counts. They give 8 content domains as performance score BANDS on the 160-760 scale plus a question-count range ("Information and Ideas / (26% of section, 12-14 questions) / Performance: 610-670"). The existing patterns hunted for "Skill 7/12", which does not appear anywhere on an official report, so every CB upload fell through to manual entry. The KD College Prep handler worked because that provider does print counts. Now parsed: 8 domains → 9 of our skills, total/RW/Math scores, and the test date. Two traps found: CB hyphenates "Problem-Solving" and our seed data doesn't (silently dropped that whole domain), and a top band over 4-6 questions rounds to 5/5 which reads as "you got everything right" when the report says no such thing — clamped, and the review screen now says CB figures are estimated from bands rather than counted. Saravanan's out-of-order point was a live bug: get_score_report ordered by uploaded_at DESC and compute_score_trend took reports[0]/reports[-1] from upload order, so uploading an older report later made it the "latest" and produced a confident fictitious score drop. Now ordered by coalesce(test_date, uploaded_at). Only fixable because the parser can read the test date. Also: rw_score/math_score/total_score have existed since migration 030 but nothing ever wrote them — Home's report card rendered "—" for every score. Vision (#148) is built around refusing rather than guessing, because extracted numbers land in user_skill_state and drive question selection: the prompt offers {"readable": false} as a correct answer, out-of-range values are dropped not clamped, RW+Math must equal Total or all three are discarded, dates outside 2023-today are rejected, under four readable domains sinks the extraction, temperature=0. Most of the 25 new tests assert it returns nothing. A photographed report and a PDF produce byte-identical rows — both run through cb_domains_to_entries, with a test pinning it. gemma4 chosen: cheapest vision model in Ollama's cloud catalogue ($0.14/$0.40 per 1M) and it does text too, so one model string also switched on rationale generation (#82), which had been wired but off since the 24th. Judge stays on o4-mini — the whole point of per-role config was never moving the quality gate on cost. 202 backend tests pass (45 new); the single failure is the pre-existing test_question_validation one. NOT verified: no call has ever hit the real Ollama provider — every vision test mocks the client, so the first real call is the first image upload. Extraction accuracy is unmeasured (REQ-23.13 wants 90%); one sample report, no labelled set. · saravanan@scrumclaw.ai

activeUX review P1+P2 complete in the working tree — STAGED, NOT YET DEPLOYED. Saravanan has to run ./scripts/ship.sh; I have no gh/git credentials in this environment. 16 of 16 review items now done or explicitly parked. New: Home focus view replacing the widget Dashboard (latest report → top weak skills → one-line week strip), with the no-report state turning the whole page into the upload prompt; /upload as a real front door with a mandatory review-and-correct step; nav cut from nine items to four (Home/Practice/Progress/Full Tests) with a new /full-tests hub; Mistakes/Saved/Vocab folded into Practice as tabs and Practice defaulting to weakest-first; "Needs work / Getting there / Solid" replacing bare percentages on web and mobile; practice wrap now shows the mastery move and a "Practice 10 more"; countdown down to days, window toggle hidden until 3+ days of data, Pro badges hidden from subscribers; parent dashboard gets a read-only "where X stands" summary at the top. Two defects found while implementing, both pre-existing: (1) the Planner upload flow auto-saved whenever the parser claimed success, so a mis-read row went silently into user_skill_state and from there into weak-skill detection and question selection — /upload now never writes without confirmation; (2) parse-score-report returned 400 "File must be a .pdf", making BRD #23's own premise (photograph your score report) a dead end — images now accepted and answered with success:false plus a usable message. Verification: web + mobile typecheck clean, 165 backend tests pass (1 pre-existing failure in test_question_validation unrelated to this work), 22 new backend tests, and a link audit confirming every internal to=/href= target resolves to a declared route. vite build unverified — rollup's native binary in node_modules is darwin-only and I run on Linux, so CI is the first real build. Three decisions waiting, all written up in docs/UX_REVIEW_2026_10_24.md: can a parent upload on a student's behalf (from-score-report is require_student, so the parent summary ships read-only); mobile bottom tabs left at five rather than collapsing Exam+Predicted, because that would reverse #77; and when /dashboard gets retired. Mobile parity for Home and upload is still #87. · saravanan@scrumclaw.ai

activeUX review P0 block shipped and verified live on adaptivesat.ai. Six fixes from the 24 Oct UX review: (1) /history 404 now redirects to /exam/history; (2) removed "this will replace your current score report" — migration 030 made reports a history table, copy now says every report is kept; (3) /practice?skill=<id> routes straight into the drill instead of dumping users on the section picker; (4) the "take a full exam to see weak skills" prompt is now gated on getScoreReportStatus() so it hides once a report exists; (5) "Upload score report" buttons added to Dashboard and Progress (upload was previously reachable only via Planner); (6) every skill row on Progress is now a link into /drill/:skillId with a hover "Practice →" affordance. Verified against the deployed bundle assets/index-DAHnVl2K.js — /drill/ x3, "Upload score report" x2, new copy x1, old contradictory copy x0, /exam/history x3. Backlog tracked in docs/UX_REVIEW_2026_10_24.md with all 16 review items and status. Next: P1 is the Home focus view + nav reduction to four items (items 7-11), and one open decision first — "Drill" vs "Practice" as the single verb (item 13), cheap to settle now at one route and two components. · saravanan@scrumclaw.ai

activeMicro-training drill (BRD #23 REQ-23.7/7b/7c) shipped and verified live — api.adaptivesat.ai now serves 119 paths including all four /api/drill endpoints (start, next, answer, end). A drill is a PRACTICE session with session_type='drill' and a preselected blueprint over one weak skill, reusing the existing session/attempt plumbing rather than a parallel one. Each answer returns the four-part feedback block from the sample tutoring session: verdict + running score, key clue quoted from the underline span, per-distractor rationale (student's own wrong choice first), and a gap note. Adaptive length as decided: stops at 3 correct in a row with a floor of 3 questions, caps at 10. Supersedes the fixed 12 from the sample — 10-15 min is the length already failing on mobile, and ending once a student has demonstrably got it is better evidence than making them grind the rest. Gap tracking is DERIVED from the session's attempts, not stored: a subtype opens on a miss, closes on a later correct, reopens on regression. Deliberately not persisted — the attempts table already knows, and a second copy is how the mastery_score defect happened. Two places it refuses to guess: _key_clue returns None when there's no underline span rather than inventing evidence, and a rationale-provider failure degrades to the single explanation rather than breaking the drill. 24 new tests, 127 passing. NOT YET USER-REACHABLE — API only, no UI on web or mobile. Also none of it has run against a real database or a real question; the tests are pure-logic with mocks, so first real call is the actual test. · saravanan@scrumclaw.ai

activePRs #134-#136 shipped: documentation brought back in line with reality. The implementation status doc was materially wrong, not just stale — #8 Subscription said "Not started" when Stripe web billing has been live since 2026-06-16, and #5 iOS said "Active focus, EAS Build is the gating step" when it has actually been rejected twice and sitting unsubmitted for three months. Both corrected. Added BRD #23 to the status doc with shipped-vs-remaining. Backfilled the commit log, which had stopped at 2026-04-30 — 82 commits from May through October restored from git, now 133 total. BRD_V3_TRACKER REQ-19.5/19.7 annotated as partly superseded by migration 030. BRD23 doc now carries a per-REQ build status table. Headline risk, now written down in the status doc: #5 is the only thing in this project that has not moved. Build 7 was produced 2026-09-08 and never submitted; it now predates four merged PRs (#124 nav, #126 exam exit, #132/#133 score history), so a fresh build is required regardless. The App Review notes on the version also still quote $0.99/$9.99, which will contradict the live price once the pricing change is configured — a rejection risk on its own. d05ff09 · saravanan@scrumclaw.ai

activePR #133 shipped: score history is now visible end to end. GET /api/study-plan/score-history returns every uploaded report plus baseline-vs-latest per-skill deltas; parent mirror at GET /api/parent/students/{id}/score-history (REQ-23.10, read-only, guardianship-gated). ScoreHistory component on web Planner + Progress and mobile PlannerScreen + ProgressScreen. 10 new tests lock the semantics: trend is baseline→latest not latest-vs-previous (20%→90%→50% reads +30, not -40); a skill only in the newest report goes to new_since_baseline rather than being reported as a swing from zero; unmapped CB labels still trend by normalised raw label. Docs updated — BRD23 now carries a per-REQ build status table, BRD_V3_TRACKER REQ-19.5/19.7 annotated as partly superseded by migration 030. Earlier in the session: frontend deploys had been failing silently since 21 Oct (firebase-tools@latest added a Service Usage precheck, github-deployer lacked serviceusage.services.get) — granted the role, pinned the CLI, and both deploy workflows now open a GitHub issue on failure. Prajith's Planner nav tab is finally live. 9b98ed8 · saravanan@scrumclaw.ai

activeBRD #23 (score-report upload → baseline → micro-training) scoped and foundations shipped. Fixed #79: mastery_score normalised to 0-100 percent (migration 027, recomputed from attempts + CHECK constraint) — the weak-skill detector was comparing an integer column against 0.7, so it returned every untouched skill instead of genuinely weak ones; 157 of 176 rows were seeded zeros. Also now excludes never-attempted skills via last_seen_at. Leverage/Today's Plan verified working in prod. Added cached per-distractor rationales (migration 028) on a swappable LLM role so high-volume rationale work can run off OpenAI without moving the judge. Reconciled a collision: Prajith and I both built score_reports on 2026-10-21 to incompatible designs — his shipped first as one-row-per-user, migration 030 converts it in place to a history table (drops the unique constraint, moves entries JSONB into score_report_entries, drops the blob) since history is the point of #23. Pricing moving to $1.99/mo + $19.99/yr. Root-caused a 2-day silent frontend outage: firebase-tools@latest added a Service Usage precheck that 403'd because github-deployer lacked serviceusage.services.get — Prajith's Planner nav tab has been invisible in prod since 21 Oct. Pinned the CLI and added failure-reporting so a broken deploy opens a GitHub issue instead of going unnoticed (third silent deploy failure this quarter). · saravanan@scrumclaw.ai

blockedPlanner changes shipped via PR (nav tab, 'score report uploaded' view with replace option, upload redirects to Progress). Backend deploy is green: migration 029 (score_reports table) applied and GET /api/study-plan/score-report is live. Frontend deploy FAILED (run 39710509068): 403 permission denied on firebasehosting.googleapis.com for the GitHub Actions service account, so the new UI is not on adaptivesat.ai yet. Tracked as a blocker. NEXT: grant the GHA service account Firebase Hosting Admin (and Service Usage Consumer) on prediction-383914, then 'gh run rerun 39710509068 --failed', hard-refresh the site and test the Planner tab with the KD PDF. Note: accounts that uploaded a report before 029 have no stored report record and must re-upload once. · prajith.k19022003@gmail.com

activeScore report upload fixed for KD College Prep reports and verified working on the live site. Cause: the parser in planner_service.py only matched College Board-style 'x/y' rows, but KD reports list '<type> (<ABBR>) <correct> <missed>'. Added _extract_kd_rows: reads the abbreviation, rolls up into the 11 domain-level skills (totals match the report's own domain percentages), and falls back to the old patterns. Tested on a real KD PDF (9 rows found, 64/98 correct) and shipped via PR by owner. Limits: only KD and College Board-style layouts are recognized; the History/Social Studies and Science analysis skills get no score from KD reports; no unit test yet. NEXT: fix web navigation issues (owner to list them). · prajith.k19022003@gmail.com

activeStatus refresh: 10 open / 11 done, no blockers, no due dates. Supersedes update #20 (2026-07-23, "iOS Waiting for Review") — iOS was rejected a 2nd time on 2026-07-24, tracked in #38. Critical path: #38 iOS build 7 resubmit → #71 Bluebook nav → #39 mobile highlighting. In parallel: #36 Android/Play, #17 audit cron, #35 tech debt. · saravanan@scrumclaw.ai

activeDocs now PUSHED (updates prior status): owner ran ship.sh → PR #113 (d437cd8) merged to main — docs/BRD5_RESUBMISSION_2026_07_23.md + .gitignore asc_submission/. Working tree clean. Remaining from tonight: only #17 (REQ-13.11 cron 7 clean runs) — verify with the gcloud command in the comment on #17, close if clean. iOS submission still Waiting for Review (Apple email ~48h, auto-release on approval). d437cd8 · saravanan@scrumclaw.ai

activeiOS resubmission SHIPPED 2026-07-23: one App Review submission bundling app v1.0 build 6 + Premium Annual subscription + AdaptiveSAT Premium subscription group — status Waiting for Review (~48h, auto-release on approval). Rejection fixes: SIWA (4.8), age rating Parental Controls→None=4+ (2.3.6), account-delete + screen recording (5.1.1v). Closed tracker #22 (resubmission) and #34 (native IAP). Documented in docs/BRD5_RESUBMISSION_2026_07_23.md (+ .gitignore asc_submission/); NOT yet pushed — no git creds in sandbox, owner runs `rm -f .git/index.lock && ./scripts/ship.sh "..."`. #17 (REQ-13.11 cron 7 clean runs) could NOT be verified from sandbox (no gcloud) — see comment on #17. · saravanan@scrumclaw.ai

activeSession wrap 2026-07-01 — iOS monetization + Apple auth + a production exam bug. IOS IAP (BRD #8 follow-on) — fully wired end-to-end (config, not just code): • App Store Connect: created subscription group "AdaptiveSAT Premium" + auto-renewable sub product `com.adaptivesat.premium.annual`, $9.99/yr across 175 countries, English localization, and a 7-day free-trial introductory offer (Jun 2026–Jun 2027). Paid Apps agreement confirmed Active. (Review screenshot still needed at submission time.) • RevenueCat: project + App Store app (bundle com.hopewellpartners.adaptivesat, .p8 IAP key uploaded by owner), entitlement `premium`, product, default offering ($rc_annual → App Store product), webhook → https://api.adaptivesat.ai/api/billing/revenuecat/webhook (auth secret in GCP `adaptive-sat-revenuecat-webhook-auth`, wired into deploy-backend.yml; webhook verified 401→live). • Repo: appl_ SDK key in eas.json (PR #100); ascAppId corrected 667811167→6766082796 (was the team id — would’ve broken eas submit). Kept the existing custom MobilePaywall (chose not to adopt RevenueCat prebuilt Paywalls/Customer Center). Decision recorded: the pasted “integrate SDK” prompt was RevenueCat’s generic one — SDK already integrated; ignored its test_ key + placeholder ids. Sign in with Apple — was missing on the new backend after migration: • iOS native fixed: APPLE_OAUTH_AUDIENCES=com.hopewellpartners.adaptivesat added to deploy-backend.yml env. • Web SIWA enabled (owner wanted it — Mac/US users): created Apple Services ID com.hopewellpartners.adaptivesat.web (domain adaptivesat.ai + return URL /auth/apple/callback), frontend VITE_APPLE_SERVICE_ID set (PR #103), backend audience appended (…,…adaptivesat.web). No SIWA private key needed (client id_token flow verified against Apple JWKS). Predicted Tests discoverability (real gap the owner caught): the Premium paper hub had NO UI entry point on web or mobile (URL-only). Fixed: web top-nav + mobile-web menu link (PR #105); iOS Dashboard card (PR #106, ships next build). PROD BUG — “Failed to start exam” (browser 503 → actually 500): root cause = OpenAI account out of quota (429 insufficient_quota) during on-the-fly question generation in exam start. NOT migration-related. Fixes: PR #107 (exam-start survives generation failures → falls back to existing bank questions, clean 400 instead of 500), #108 (email the owner on quota exhaustion, throttled hourly; openai_alert_email default saravanan.hp@gmail.com / OPENAI_ALERT_EMAIL), #109 (missing `import time` that re-crashed it), #110 (quota cool-down so it fast-fails instead of hanging ~60s). VERIFIED: exam/start now returns 200 (session created) building entirely from the existing bank — OpenAI credits now OPTIONAL (only needed for fresh/no-repeat generation). Owner got the alert email. Also earlier this session: retired old backend (deleted adaptive-sat-api in prj-crypto538); DB pw rotation deferred (shared role); tracker items #37 (migration) + #18 (REQ-13.9) closed; REQ-13.11 cron clock running (reminder set for Jul 4). OPEN / NEXT: (1) Task “iOS sandbox purchase test” — owner needs a fresh EAS preview build (has appl_ key + Predicted Tests card) + a Free test account (early-member accounts bypass the paywall) + sandbox Apple ID. (2) ASC subscription review screenshot before submission. (3) Top up OpenAI whenever (optional). (4) Rotate the migration GitHub PAT. (5) Android Play Store, iOS resubmission still pending per prior items. · saravanan@scrumclaw.ai

activeSession wrap 2026-06-27. Closed the accuracy phase. (1) REQ-13.11 nightly quality cron DEPLOYED in prediction-383914 → api.adaptivesat.ai (both Scheduler jobs live + verified 2xx; 7-day clock running, closes ~2026-07-04; reminder scheduled for July 4 to verify + flip to Done). (2) REQ-13.9 DONE (item #18 closed): nightly audit re-judged all 48 served questions through the fact_safe gate — {"sampled":48,"failed_judge":0,"queued":0} = zero named-entity hallucinations; owner accepted the served-only sample (306 live-but-unserved questions get audited as they accrue attempts). Docs corrected post-migration: cron runbook (PR #97), BRD_V3_TRACKER (PR #98), IAP runbook (PR #99). NEXT (owner-gated, paused by request): mobile monetization — #34 iOS IAP config (ASC sub + .p8 key, RevenueCat project/entitlement 'premium'/offering/webhook, then wire REVENUECAT_WEBHOOK_AUTH secret line + EXPO_PUBLIC_REVENUECAT_IOS_KEY in eas.json) → #22 iOS resubmission → #36 Android. All code shipped/dormant; remaining work is Apple/RevenueCat/Google account actions only the owner can do. · saravanan@scrumclaw.ai

activeREQ-13.11 nightly question-quality cron DEPLOYED (2026-06-27). Both Cloud Scheduler jobs created in the NEW project prediction-383914 pointing at api.adaptivesat.ai: question-wrong-rate-flag (daily 10:00 UTC) and question-nightly-audit (nightly 09:00 UTC). Enabled cloudscheduler.googleapis.com in the new project; admin key confirmed present. Verified: wrong-rate manual run = 200; nightly-audit first scheduled run (09:00 UTC) returned 2xx (status {}). The 7-consecutive-clean-runs acceptance clock for REQ-13.11 starts now — closes ~2026-07-04 if no errors. Runbook corrected to the new project + stable domain (PR #97). Remaining on item #17: confirm 7 consecutive clean runs, then mark REQ-13.11 Done. · saravanan@scrumclaw.ai

activeGCP project migration fully closed out. AdaptiveSAT now runs entirely on its own project prediction-383914: backend on Cloud Run behind stable api.adaptivesat.ai, frontend on the new Firebase site serving adaptivesat.ai, CI (deploy-backend/frontend) retargeted to the new project. DB kept in place in prj-crypto538 and reached cross-project (cloudsql.client grant + --add-cloudsql-instances). Stripe webhook repointed to api.adaptivesat.ai AND the pre-existing 500 bug fixed (PR #90: process raw JSON dict since stripe-python StripeObject dropped dict .get(); read current_period_end from items fallback) — subscription rows now write; test user confirmed status=active. Mobile config + eas.json on the stable domain (PR #95). Old Cloud Run service adaptive-sat-api deleted from prj-crypto538; api.adaptivesat.ai still 200 after. DB password rotation DEFERRED by decision (shared adaptive_app role). Only loose end: revoke the migration GitHub PAT. Checklist closed out in docs/MIGRATION_CUTOVER_CHECKLIST.md (PR #96). · saravanan@scrumclaw.ai

activeNative iOS IAP (RevenueCat) build SHIPPED + dormant. Code merged to main via ship.sh; migration 025 applied in prod (Run-migrations workflow, success); webhook route live and returns 503 until configured (correct dormant state). Predicted Tests Pro routing finished (REQ-17.3 Done): web Paywall, iOS native RevenueCat paywall, Android web-fallback. Docs updated: BRD8_TRACKER (operational done + IAP follow-on), BRD8_SUBSCRIPTION_SPEC §17, BRD_V3_TRACKER (REQ-17.3 + changelog), IAP_SETUP_RUNBOOK.md. Remaining (owner): App Store Connect product + RevenueCat config + GCP secret + deploy-backend.yml line + EAS build + sandbox test (item #34). · saravanan@scrumclaw.ai

idleSession wrap. Done this session: (1) math-in-reading bug fixed + shipped (PR #83) — exam-header section-label desync, not a data bug; (2) BRD #8 Subscriptions fully LIVE — Stripe setup, GCP secrets, Bundle F (PR #84), passing trial-signup smoke test; (3) seeded 20 hand-authored+self-judged d5 questions (10 Math/10 R&W) APPROVED via scripts/seed_d5_claude_questions.sql (tag source_model='claude-d5-v1'). Open follow-ups: in-app symbol-render spot-check of the d5 batch; optional save as JSON seeds. Top remaining items: #17 nightly quality cron, #22 iOS resubmission, #19 Predicted Tests Pro routing. · saravanan@scrumclaw.ai

activeSession wrap: (1) Math-in-reading bug fixed + shipped (PR #83) — was an exam-header section-label desync, not a data bug. (2) BRD #8 Subscriptions fully LIVE — Stripe setup, GCP secrets, Bundle F deploy (PR #84), live trial-signup smoke test passed (7-day free trial before first charge). Items #14/#15/#16 closed. User chose to let the test subscription's first charge happen in ~7 days and sort it out then. · saravanan@scrumclaw.ai

doneBRD #8 Subscriptions COMPLETE — live checkout smoke test passed end-to-end (signup → Stripe Checkout → payment → webhook). All three go-live items (#14 Stripe setup, #15 GCP secrets + Bundle F deploy, #16 verify/live) closed. Billing live in production. · saravanan@scrumclaw.ai

activeBRD #8 billing LIVE. Stripe dashboard (Live mode) set up: product AdaptiveSAT Premium $9.99/yr, Tax on (origin Frisco TX), Customer Portal active, webhook with 6 events. 3 Stripe secrets in GCP Secret Manager + IAM. Bundle F (deploy-backend.yml secrets wiring) merged via PR #84, backend deployed. Webhook endpoint returns 400 (signature verify active) = endpoints off 503. Remaining: end-to-end live checkout smoke test (item #16). · saravanan@scrumclaw.ai

idleProject initialized: Adaptive SAT · saravanan@scrumclaw.ai