[coderconnoisseur]

UI revamp Phase 3: the backend behind Home. Stacked on #92 → #91 → #89. Merge them in that order and retarget each PR to refactor as its parent lands.
Spec: docs/superpowers/specs/2026-10-20-ui-revamp-design.md §4, §8. Plan: docs/superpowers/plans/2026-10-21-ui-revamp-phase3-home-backend.md.

Backend

• GET /api/home/summary (api/home.py over the pure ai/home_signals.py, no LLM):
  • Lumen found:
    • Possible double charges, the only items with a "you may get back" amount. A vendor you pay most days (3+ charges in 30 days) is a habit, not a duplicate.
    • Receipts to review.
    • Category spikes: ≥30% up on the same day last month and at least ₹1,000 / 5% of last month. Spending only, so refunds can't fake one.
    • Subscription price rises (≥5%, from ai.recurring).
  • Activity:
    • Statement imports, grouped into one line per import.
    • Receipts read.
    • Your approve/reject decisions.
  • Text carries {0} placeholders plus numeric amounts, so every amount shows in the user's currency.
• Budgets:
  • The new budgets table holds per-category limits. The overall budget is still the profile's monthly_budget.
  • GET/PUT /api/budget.
  • POST /api/budget/draft runs a deterministic draft (mean of the last 3 complete months that had spend, rounded up; needs 2+ months) and files one set_budget proposal per category. Approving it through /api/agent/proposals/{id}/approve upserts the limit, and only a finite positive limit is ever applied. A new draft retires pending ones.
  • Ask Lumen gets a draft_budget tool that files the same approvable proposals.
• GET /api/review?count=1 returns just the number, for the sidebar badge.

Frontend

• Lumen found: up to 3 tiles from the summary, with "N things worth a look · ₹X you may get back".
• Brief: adds "The one thing worth a look: …".
• Budget card: progress against the monthly budget (left, days, per-day, or "over"), plus category rows.
• "Draft budget": opens a dialog to approve lines one by one or all at once. It is guarded against double clicks.
• Activity: comes from the server.
• Fixes from the backlog: the calendar uses the user's currency symbol, and a refund-heavy month can't flip the brief.
• Sidebar badge: uses the count endpoint.

Checks

• Backend: pytest -q gives 950 passed, 15 skipped. That includes 47 new tests:
  • draft rounding and history rules
  • per-user isolation
  • a double approve returning 409 with one row
  • re-drafts retiring stale proposals
  • bad agent payloads returning 422
  • refunds and habits not producing signals
  • UTC Z timestamps
• Frontend: npm run test:e2e gives 82/82. tsc, lint and build are clean, and /home is static at 17.5 kB of route JS.

Review (whole branch, Opus)

• Fixed with tests:
  • An agent-filed set_budget with a bad limit could break /api/budget.
  • Drafts built from 2 months of history divided by 3.
  • Refunds faked spikes.
  • A daily coffee showed up as "double charges".
  • The draft dialog double-submitted.
• Deferred minors are in spec §12:
  • an approval race without a conditional update
  • drafts retiring each other across Home and Ask
  • Activity not refreshing after approvals
  • full-scan cost per visit
  • per-line button names
• Owner request, logged in §12: make Lumen found and budget drafts smarter about one-off big purchases (median or trimmed mean, planned purchases), measured against an eval set.

Manual check

• On a demo account: Draft budget → approve all → the Budget card shows the category rows; Lumen found makes sense against the data.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
lumen	Ready	Preview	Oct 21, 2026 5:32pm UTC