accor-activity

Unified output contract. This sub-skill is part of points-activity. Its scripts/transform.py writes Date, Description, Amount and the filename accor_activity_<from>_<to>.csv (covered range) via the shared scripts/activity_output.py, and prints BALANCE:/COVERED:/REQUESTED:/FILE:/ROWS: to stdout. The orchestrator passes the requested period and reads those lines. References to collapsed.csv / Points / Miles below are the older single-program wording; the actual columns are Date, Description, Amount.

Accor ALL Points Statement Extraction

Scrapes the rendered Accor points statement (global transaction history), expands all year accordions, keeps only REWARD points, collapses, and outputs a points table.

Prerequisites

• Claude in Chrome extension connected (list_connected_browsers → select_browser).
• User logged in to all.accor.com. Don't log in on their behalf.
• Verify login programmatically: the page title is "Your points statement ..." and a member name / reward-points balance is shown; there are no sign-in buttons.

Why DOM scraping (not the API)

Accor's data sits behind api.accor.com with a Bearer JWT (localStorage['identification-token_all.accor']) plus apiKey + x-caller headers. The auth can be reconstructed (and e.g. /orders/v1/contacts/me/orders returns 200 — but only bookings, 2 of them, not the point lines). The actual points-statement endpoint never shows up in network logs or fetch/XHR hooks — the data is delivered at first render and cached in Angular state. The rendered table is complete and its reward-points sum reconciles to the balance, so scraping wins.

Procedure

1. Open the page

tabs_context_mcp(createIfEmpty:true) → navigate to https://all.accor.com/account/en/global-transaction-history. Confirm login. Keep tabId.

2. Scrape (expands years automatically)

Run scripts/scrape_activity.js. History is grouped into year accordions; the current year is expanded, older years are COLLAPSED and their rows aren't in the DOM until the year header is clicked. The script clicks every collapsed year, then parses each <tr class="simple-grid__tr--tbody"> (5 cells: Description, Date MM/DD/YYYY, Reward, Status, Nights), keeping year context. Returns {count, years}, stashes rows on window.__accor.

3. Dump and read back

Run scripts/dump_console.js (logs AC###~ ...), then read_console_messages(pattern:"^AC\\d", limit:60, tabId:...). (Direct JSON returns are often blocked by the chat — the console path avoids it.)

4. Transform

Strip the AC###: prefix, write the ~~-delimited lines to a file (e.g. /tmp/raw.txt), then:

python3 scripts/transform.py /tmp/raw.txt <output_dir> [from] [to] [balance]
# from/to are ISO yyyy-mm-dd or '-' for unbounded; balance is the current balance or omit.

Writes collapsed.csv, prints row count + reward total.

5. Present and sanity-check

present_files on collapsed.csv. As a sanity check, the reward-points total may equal the current reward-points balance (true for accounts whose full history is visible). If it does, great; if not, don't force it — older history may have rolled off.

The collapsed table (the deliverable)

collapsed.csv: columns Date, Description, Points (Points = REWARD points only; Status points and Nights are intentionally ignored). Logic (in scripts/transform.py):

1. Classify: redemption = description starts "Stay paid with points:"; everything else (stays, Platinum bonus, partner earns like Flying Blue / Europcar / BILT) is an earning/transfer.
2. Redemptions: keep the real transaction date; each stays its own row (matched to invoices later); hotel name kept in the description.
3. Earnings/transfers: move the date to the last day of its month so repeated monthly items collapse.
4. Collapse by (Date, Description), sum reward Points.
5. Drop zero-point rows after collapsing (rows that only carried Status/Nights, or that net to zero, fall out).

Recurring / scheduled use (monthly)

Run as a scheduled task on the user's own machine — needs the user's real Chrome + live Accor session. Use the host's scheduling primitive: Cowork — mcp__scheduled-tasks__create_scheduled_task; Claude Code — cron + headless invocation. Caveats: local tasks run only while the computer is on and the app is open; no completion/failure notifications yet; an expired session needs a fresh login first. Each run is a clean session — to report only new activity, save each run's collapsed.csv (dated filename) and diff against the previous one.