Backlog

Fetch statement_balance, next_payment_due_date, minimum_payment_amount, last_payment_amount, apr per card. Foundation for F1.2-F1.7.

Split "$7,021" into "Due May 5: $X" + "Accumulating for next cycle: $Y". Same total subtracted; clarity goes up.

Push N days before each card's due date showing minimum-required amount.

"Pay at least $50 to avoid late fees" — non-judgmental safety net.

Surface APR cost in dollar terms when balance lingers. Direct savings impact.

Drag/slider showing 25%/50%/75%/100% payment scenarios with interest impact. Sticky, share-worthy.

When last_payment_amount = 0 for closed cycle, surface as alert.

Near-real-time spend tracking. Closes the post-swipe latency gap.

Surface "Pending: $X" distinctly from cleared.

Project "remaining today" instead of "fresh today".

365-day scan on every Home open is fine at dev scale, p95 regressor at 10K+ users. Schema ready.

Current pass rate ~70%. Dedicated session: teach Vik the eval system; identify failing fixtures; fix prompts/tools/grounding.

Replace custom HTTP tool-use loop with managed agent. Less maintenance.

Coach now returns this field; we have zero fixtures testing it.

Moot — Smart Pulse removed 2026-06-12; any revival rides the Epic 13 full rebuild.

Currently MVP+. Anomaly + trend analysis on transaction history.

Written in evals/schedule/ (2026-10-04) but not switched on: ~$20/month on AWS. Load the two launchd files to enable. (Evals Phase 7, deferred 2026-10-04.)

~100 owner yes/no labels per judged type (ambiguity, distress, premise, freshness, stance), split 20/40/40, test once. Until then judge results are provisional. Plan: evals/judges/per_type_judges.md. (Evals Phase 7, deferred 2026-10-04.)

The corrected-rate tool exists (npm run corrected-rate); needs certified judges (F4.7). Then retire the old 92-test scripted suite. (Evals Phase 7, deferred 2026-10-04.)

Keeps the failure-type map current (round 2 found two new types) and measures grader agreement. (Evals Phase 7, deferred 2026-10-04.)

chat/[id].tsx never passes messageId and ai_traces.message_id is never written, so feedback is lost. (Evals Phase 7, deferred 2026-10-04.)

Frozen replays + the same judge make a model switch a measured decision. Also: one-time Bedrock vs Anthropic API parity check (needs an API key). (Evals Phase 7, deferred 2026-10-04.)

Round 2 open product question behind most remaining overcautious Coach answers (failure type 16). (Evals Phase 7, deferred 2026-10-04.)

~13% of transactions don't have Plaid category. Haiku categorizer fills these in.

User corrections in merchant_category_overrides feed back into AI prompts. Compounds.

Backend exists; UI is MVP+. User taps transaction → "change category" → applies to all future txns from that merchant.

Currently MVP+, dev-bypassed during sprint. Replaces dev-bypass.

Sequential Plaid Link with "add another bank" CTA.

Handle ITEM_LOGIN_REQUIRED webhook + reconnect prompt in Settings.

Extend "we'll show your pattern after a few weeks" pattern to all of Safe to Spend.

Hook up event triggers; foundation for F7.2, F7.3, F1.3, F14.2.

Moot — Smart Pulse removed 2026-06-12; any revival rides the Epic 13 full rebuild.

"Your $1,200 rent hits Friday — do you have enough?"

Daily reminder at 8pm; tag transactions "worth it" / "regret"; completion tracked via reflection_sessions (streaks removed 2026-06-11).

Long-press transaction → mark; feeds query_reflection_stats worth-it ratio.

Tool removed 2026-10-04 (round-1 error analysis, failure type 14): with the reflection UI gone it always returned empty while the prompt told the Coach to cite reflection tags. Rebuild only together with F8.1.

"Save $5K by August"; goal thermometer.

Plaid + linked savings account integration. Set-and-forget, highest impact.

"Round Starbucks to nearest dollar; send the change to savings."

When income > expected, suggest saving the excess.

"You skipped lunch out 3× this week → $45 saved → move to savings?"

Goal thermometer + projected hit-date.

Consistency gamification.

Full rebuild — the entire challenges implementation (mobile Progress tab, backend controllers/jobs/policies, 5 DB tables) was removed 2026-06-10 ("less is more" cleanup; recover via git history). NOT a re-enable.

"You spent $400 dining last month — try $300?" Depends on the F9.8 rebuild; the old challenges code no longer exists in the tree.

Needs schema (per-occurrence is_paid flag).

Currently Sandbox/Development. Production requires Transactions product approval (in flight).

iOS small + iOS Pro Max + Android. Surface UI bugs.

Hot paths (typical-discretionary scan, recurrence detection, AI tool calls) need profiling. RDS sizing review.

End-to-end pass through every doc, every GitHub Pages visualization, every LEARNINGS.md entry, every README — verify each is current, comprehensive, and accurate. Clean up GitHub: stale branches, orphan PRs, untriaged issues. Tidy local Git: drop ignored junk (.expo/, .obsidian/, etc.), confirm .gitignore is right. The bar: anyone reading the docs can understand the app without needing to ask Vik or Claude.

Currently fixed 7-day; adjust to user's payday cadence.

Full rebuild — the hidden Pulse History screen, its Home-header flame/streak-pill entry, the GET /morning-briefs list/show endpoints, AND (same day, follow-up PR) the entire backend streak feature — StreakService, /streaks endpoints, user_streaks + streak_milestones tables (dropped by migration 1776000000017) — were removed 2026-06-11 ("less is more" cleanup; recover via git history). NOT a re-enable — morning_briefs rows still accrue daily so brief data survives, but streak history does NOT (tables dropped).

Adapt to time of day, recent activity, day-of-week patterns.

"You could save $80/mo by paying $1K extra to Chase" → tap → CC payment / transfer screen.

Morning brief 8am, midday nudge if overspent, evening reflection prompt.

Tap Pulse → see prompt context (transparency).

Separate fixtures from Coach evals; pulse prompts have their own quality bar.

Fall through to deterministic rule when AI is uncertain.

Feed signal back into prompt tuning.

Detect when (current runway − sum-of-bills-due-in-next-N-days) < threshold. Configurable per-user. Runs in generate_predictions_job.ts.

Fires when shortfall detected; deep-links into a detail screen. Depends on F7.1.

List upcoming bills, show the gap, suggest options: "transfer $X from savings", "delay this bill", "you typically spend $Y/day — cut back N days?"

Let user mark a bill "I'll pay this on [later date]"; updates runway projection.

Partnership with cash-advance product (Earnin, Brigit). Out of scope unless we hit a revenue model that justifies it.

Mark a transaction as "one-time, exclude from averages." Inverse of F2.1. Closes a known typical-discretionary distortion.

"Treat this checking as savings — don't count toward Safe to Spend." User control over v4.9 setAsideAccounts boundary.

Differentiate Spotify (cancellable) from rent (committed). Powers F14.3 "what could you cut?" suggestions.

In-app feedback on each Coach response, fed into eval fixture pipeline. Compounds with F4.1.

"This isn't 'Amazon' — it's a local salon called 'Amazon Salon'." Powers a global merchant-alias map.

Let user input expected paycheck amount when only one observation has happened. Addresses v4.5 "need 2+ observations" gap.

One place to see all your overrides + confidence flags + correction history. Transparency win.

Automated mechanism to roll user corrections into prompt context for next AI call. The meta-system that makes every other Epic 15 item compound. The AI moat.

The merchant normalizer mints a new merchant for nearly every transfer row, so grouping, recurrence detection and pending/posted matching miss repeats. Found by the real-data evals (2026-10-04); the Coach groups transfers by normalized description as a workaround.

Read DESIGN_SYSTEM.md, gap-check against actual code. Produce a delta report: documented vs hardcoded vs missing.

Automated rule: fail CI when hardcoded colors/spacing/typography appear outside tokens.ts. Blocks regression.

Catalog every UI component, document props/variants/states/token usage. Storybook or simpler MD table.

Expand: tokens, components, spacing, typography, motion, haptics, accessibility, layout grids.

Discovery: what's good, what's basic, what's competitive, what's the brand we want to project. Possibly paid consultant.

Opinionated transitions, micro-interactions, haptic patterns (success haptic on bill confirm, subtle bounce on Safe to Spend value change).

Currently afterthoughts. Make them feel intentional, on-brand, sometimes delightful.

Phone → OTP → biometric → Plaid Link should feel premium, not generic.

Current palette is functional; explore distinctive identity (think Cash App green, Mint coral).

Expressive type for hero numbers, more confident hierarchy. Maybe a custom display face.

Consistent custom icons replacing the Ionicons stock set; differentiates from "looks like every other RN app."

Promote reusable bits we've built (DrillDownRow, IOSListRow, AnimatedNumber) into a documented system.

Plaid Investments product (Fidelity, Schwab, Robinhood, Vanguard). Pull holdings, balances, performance.

Extend Plaid Liabilities to mortgages, auto loans, student loans (separate from Epic 1 CC liabilities).

Single screen: total assets − total liabilities = net worth, with breakdown.

Track growth over time; weekly/monthly snapshots in net_worth_snapshots; chart progression.

Equity / bonds / cash / real estate / crypto split visualization.

Home value (Zillow API or user-entered), vehicles (KBB), valuables. Plaid won't have these.

Structured "income changes / expense changes / one-time events" inputs; couple-tap entry.

Extends safe_to_spend_service.ts with override deltas; produces 3, 6, 12-month forward projections.

Side-by-side baseline vs scenario charts (runway, savings projection, debt paydown).

Save, name, revisit scenarios ("if I take new job", "if we have kids"); compare two saved scenarios side-by-side.

Pre-built scenarios: "raise", "lose job", "buy car", "have a kid", "move cities", "refinance mortgage".

When a Coach answer is what-if-ish, offer "want to see this projected? open scenario builder" deep-link.

Get lint/typecheck/unit green against the real backend, then gh workflow enable CI. Postgres+Redis services and full test-env wiring already exist in ci.yaml — these are real code/setup failures to fix, not infra to rebuild.

Verify the ANTHROPIC_API_KEY secret + evals/ scripts (eval:ci, eval:nightly, dashboard) run in CI; confirm the nightly dashboard auto-commit works; then re-enable. Compounds with F4.1 / F21.7.

Exercise deploy.yaml once: ECR push, k3s SSH rollout, migration step. Confirm secrets (AWS_DEPLOY_ROLE_ARN, K3S_HOST, K3S_SSH_KEY) are set. Currently never run — an unknown-good.

Require CI green before merge, once CI is reliable (F22.1). Turns the pipeline from advisory into an actual gate.