bilt-activity

Bilt Rewards Activity Extraction

Unified output contract. Part of the points-activity suite. scripts/transform.py writes Date, Description, Amount and the filename bilt_activity_<from>_<to>.csv (covered range) via the shared scripts/activity_output.py, and prints BALANCE:/COVERED:/REQUESTED:/FILE:/ROWS: to stdout.

Bilt is a bank/card program: earnings are per-purchase (hundreds–thousands of rows), redemptions are point transfers out to airline/hotel partners. We collapse earnings hard (by month + benefit item) and keep redemptions per-transaction.

Prerequisites

• Claude in Chrome connected (list_connected_browsers → select_browser).
• User logged in to bilt.com. Don't log in for them. Confirm the points balance shows in the header ("NN.Nk pts").

Extraction — JSON API (not DOM)

Endpoint: GET https://api.biltrewards.com/loyalty/activity?month=M&year=Y with header authorization: Bearer <jwt> (the JWT is in localStorage — scan for a long value starting with ey). Plain fetch(..., {credentials:'include'}) works.

Only month+year filter. limit, page, from/to, startDate/endDate are all ignored (they return the default ~12 recent rows). So iterate months backward from the current month until 3 consecutive empty months. Bilt has no history-window limit — full account history is retrievable and reconciles to the balance.

Procedure

1. Open tabs_context_mcp → navigate to https://www.bilt.com/rewards/activity. Confirm login + a points balance in the header. Decline cookie banners. Keep tabId.
2. Fetch run scripts/fetch_activity.js — finds the token, iterates months, stashes compacted entries on window.__biltall, returns {count, oldest, newest, balanceCheck}. balanceCheck should equal the displayed balance.
3. Dump & read run scripts/dump_console.js (emits BILT###~<json-array> chunks), then read_console_messages(pattern:"^BILT\\d", limit:60, tabId:...). Reassemble the chunks in order and concatenate the arrays into a single JSON array file (e.g. /tmp/raw.json). (Large dumps: read in batches; chunks are ordered by their index.)
4. Transform

   python3 scripts/transform.py /tmp/raw.json <output_dir> [from] [to] [balance]
   # from/to ISO yyyy-mm-dd or '-'; balance = header "pts" figure.
   

5. Present present_files on the CSV. Report balance + covered range.

Entry shape & the key rules

Compacted entry: {d, t (title/merchant), a (activity), s (pointState), tp (totalPoints), b: [{t (benefit item title), v (points int)}]}.

• Spendable points = totalPoints, and sum(b[].v) == totalPoints (verified). The benefit items in b are the per-line points (e.g. "2X Points on All Transactions", "Additional 1X - Point Accelerator", "3x Points on Dining"). Cash-back "Earn Bilt Cash" (+$…) and non-point benefits are dropped (the fetch script keeps only point-bearing items — those whose value parses as a number, not a $ amount).

• Classify by SIGN, not status text (robust to Bilt's vocabulary):

  • Redemption = any entry with totalPoints < 0. Transfers to partners (World of Hyatt, British Airways/Avios, Accor, Atmos, Flying Blue, …) and reversals/refunds. Keep the real date, each entry its own row, description = the title.
  • Earning = totalPoints >= 0. Explode into benefit items; move each to the last day of its month; collapse by (month, item-title) summed across all merchants — no merchant in the key. So every "3x Points on Dining" in a month becomes a single row.

• Drop zero-amount rows after collapsing (status rows like "Earned Gold status" carry no points and disappear).

Reconciliation

Full history is available, so sum(Amount) == current balance exactly (verified: 80,723 == 80.7k). If it doesn't reconcile, you likely stopped the month-iteration too early (raise the empty-month threshold) rather than hitting a history limit.

Recurring / scheduled use (monthly)

Run as a local Cowork Scheduled Task (needs the user's real Chrome + live bilt.com session). The API token expires, so a stale session needs a fresh login. Because Bilt keeps full history, a monthly cadence never misses anything; save each run's dated CSV and diff to spot new activity. For incremental runs you can fetch only the current + previous month rather than the full backfill.