Return the signed-in user's TheRealReal Obsessions (hearted/tagged items) as a structured list with current price, original price, MSRP, sale callout, designer, name, image, and product URL — keyed by internal product ID so successive snapshots can be diffed to surface price drops, new sale callouts, and delisted items. Read-only. This skill was captured from a live agent session on therealreal.com and publishes here verbatim, exactly as an agent receives it.

Purpose

Return the signed-in user's "Obsessions" — items they have hearted/tagged on TheRealReal — as a structured list with current price, designer, name, image, product URL, internal product ID, and any sale-callout metadata. Designed to be re-run on a cadence (daily/hourly) so a downstream agent can diff successive snapshots and surface price changes ("Now 20% off", "price dropped from $145 → $116", "item sold / no longer listed"). Read-only — never tap the obsession heart to un-obsess, never add to cart.

When to Use

Daily price-monitoring of a user's saved luxury items on TheRealReal.
Surfacing newly-discounted obsessions ("anything in my wishlist on sale today?").
Detecting sold/delisted items (the item no longer appears in the obsessions HTML grid).
Baseline-and-diff workflows where the agent stores yesterday's snapshot and compares.

Workflow

TheRealReal's /obsessions page is account-bound: an anonymous request to https://www.therealreal.com/obsessions returns an empty-state placeholder ("Tap the ♡ next to any item to save it for later" + "Shop by Category" carousel). To see real items the request must carry an authenticated _session_id cookie from a logged-in TheRealReal account. There is no public JSON API (/obsessions.json returns 404), so the workflow is HTML extraction from the server-rendered Rails PJAX page. The page reuses the standard Product Listing Page (PLP) markup, so every item is rendered with stable data-testid="plp-product/{productId}-{field}" selectors that survive across redesigns.

1. Acquire an authenticated session

/sign_in is hard-gated by PerimeterX's "Press & Hold" human-challenge for any automated browser session (verified iter 1 — even a stealth + residential-proxy session lands on the PressAndHold iframe). Do not attempt scripted login — it will burn the session and subsequent requests will return Access-Denied for the rest of the proxy IP's TTL. Instead, supply the user's existing _session_id cookie:

Direct cookie injection (the only path): obtain the user's _session_id value out-of-band (exported from their signed-in browser) and inject it at the start of the browserless_agent commands array with a setCookie step, before navigating (see step 2). The other required cookies (_pxhd, nearby_stores) are minted automatically on the first response. Because PerimeterX gates scripted login, there is no autonomous-login route here — the skill runs entirely off the injected session cookie.

The critical cookie name is _session_id (Rails session, HttpOnly, Secure). Without it, /obsessions returns the empty placeholder.

2. Fetch `/obsessions`

One browserless_agent call does the whole thing: inject the session cookie, navigate, and extract the cards in-page. Keep it in ONE call's commands array (the injected cookie and page state carry across the steps of the call; there's no release step, and the session isn't torn down on return — it persists across calls keyed by its config). Use a residential proxy — datacenter IPs frequently land on the PerimeterX challenge page even with valid auth cookies.

browserless_agent:

{
  "proxy": { "proxy": "residential", "proxyCountry": "us" },
  "commands": [
    {
      "method": "setCookie",
      "params": {
        "name": "_session_id",
        "value": "<USER_SESSION>",
        "domain": ".therealreal.com",
        "path": "/",
        "secure": true,
        "httpOnly": true
      }
    },
    {
      "method": "goto",
      "params": {
        "url": "https://www.therealreal.com/obsessions",
        "waitUntil": "load",
        "timeout": 45000
      }
    },
    {
      "method": "evaluate",
      "params": {
        "content": "(() => { /* parse plp-product cards — see step 4 */ })()"
      }
    }
  ]
}

The evaluate step (step 4) returns the projected item array under .value — parse the DOM in-page and return compact JSON, never ship the raw HTML back. If a call lands on the PressAndHold challenge instead of the grid, insert a { "method": "solve", "params": { "type": "captcha" } } step after the goto, or retry as a fresh call (start a distinct session so it lands on a new residential IP rather than reconnecting to the flagged one). A recent desktop User-Agent is applied by the stealth runtime automatically.

3. Detect empty vs populated

If the response contains the literal string js-empty-obsessions-message AND no data-testid="plp-product/ occurrences, the obsessions list is empty (or the session cookie was invalid/expired — there's no distinguishing signal between "logged out" and "logged in with zero obsessions" in the HTML). Return { "items": [], "total": 0, "auth_uncertain": true } and flag the user to supply a fresh _session_id cookie.

If there are plp-product/ testids in the HTML, proceed to step 4.

4. Extract each item from `data-testid="plp-product/{id}-…"` markup

Each product card is a <div role="group" data-testid="plp-product/{productId}"> with these stable child testids:

`data-testid` suffix	Meaning
`plp-product/{id}` (root)	`{id}` is the internal Rails product ID (e.g., `50758096`).
`plp-product/{id}-link`	`<a href>` → canonical product URL (relative path).
`plp-product/{id}-brand`	Designer name (e.g. "Tumi", "Louis Vuitton", "Chanel").
`plp-product/{id}-name`	Item title (e.g. "Nylon Tote", "LV Monogram Speedy 30").
`plp-product/{id}-price-msrp`	"Est. Retail $480" — the MSRP / new-retail estimate.
`plp-product/{id}-price-original`	TRR's original listing price (struck-through when on sale, else absent).
`plp-product/{id}-price-final`	Current asking price — this is the field to monitor for price changes.
`plp-product/{id}-price-callout`	Sale callout copy, e.g. "Now 20% off" (absent when not on sale).
`plp-product/{id}-obsession-button-count`	Total users obsessing this item (interest signal, not price-relevant).
`plp-product/{id}-images/image-0`	First image URL is in the wrapping `<img srcSet>`; strip `?auto=webp&…` for canonical CDN URL.

Parse the DOM in-page inside the evaluate step by selecting on the data-testid attributes — do NOT use the snapshot method to enumerate items. The accessibility tree drops the data-testid attributes and folds the visually-rich card into a single ref, making per-field extraction far more expensive than reading the live DOM by selector. Return a compact JSON projection from the eval (its return value comes back under .value); never ship the raw card HTML back.

Example evaluate content (runs in the page context — document is the loaded /obsessions DOM):

(() => {
  const items = [
    ...document.querySelectorAll('[data-testid^="plp-product/"][role="group"]'),
  ].map((root) => {
    const id = root.getAttribute("data-testid").replace("plp-product/", "");
    const pick = (suffix) => {
      const el = root.querySelector(`[data-testid="plp-product/${id}-${suffix}"]`);
      return el ? el.textContent.trim() || null : null;
    };
    const href = root
      .querySelector(`[data-testid="plp-product/${id}-link"]`)
      ?.getAttribute("href");
    const img = root
      .querySelector("img[srcSet], img[srcset]")
      ?.getAttribute("srcset")
      ?.split(",")[0]
      ?.trim()
      .split(" ")[0]
      ?.replace(/\?.*$/, "");
    return {
      product_id: id,
      designer: pick("brand"),
      name: pick("name"),
      url: href ? `https://www.therealreal.com${href}` : null,
      image: img || null,
      price_final: pick("price-final"), // "$116.00"
      price_original: pick("price-original"), // "$145" (may be null)
      price_msrp: pick("price-msrp"), // "Est. Retail $480"
      sale_callout: pick("price-callout"), // "Now 20% off" or null
      obsession_count: pick("obsession-button-count"),
      captured_at: new Date().toISOString(),
    };
  });
  return JSON.stringify({ items, total: items.length });
})();

5. Paginate (if > 1 page of obsessions)

The obsessions page is paginated. Read the data attribute data-page-number="0" off the .js-plp-data-handler element (return it from the same evaluate) and follow ?page=N links in the pagination footer. Each page is GET /obsessions?page=N with the same auth cookie. Because a long-lived session gets flagged (see gotchas), issue one browserless_agent call per page as a distinct session — same setCookie + goto + evaluate shape, just point goto at ?page=N, and don't reuse the prior call's session config — so every page fetch rides a new session on a new residential IP rather than reconnecting to the flagged one. Stop when a page returns zero plp-product/ matches.

6. (Optional) Enrich per-item via JSON-LD

If you need richer per-item data (full description, all images, US/CA condition tags, structured priceCurrency), run a browserless_agent call (residential proxy, no cookie needed — the product detail page is not auth-gated) that gotos the individual product URL, then an evaluate that parses the <script type="application/ld+json"> block whose @type == "Product" — that surfaces offers.price (numeric), priceSpecification.price (MSRP numeric), image[] array, brand.name, itemCondition, availability (InStock / OutOfStock). Only enrich items the caller actually cares about — a detail-page call per obsession balloons cost.

7. Diff against the prior snapshot

For pricing-change monitoring, persist each run keyed by product_id. On the next run, diff per-id and emit change events:

price_dropped — new price_final numerically < prior price_final.
price_raised — new price_final > prior (rare on TRR but possible after a relist).
new_callout — sale_callout appeared this run (e.g. "Now 30% off").
delisted — product_id present in prior snapshot, absent in current (sold or removed).
relisted — product_id absent in prior, present in current (unobsessed-then-re-obsessed, or formerly sold-now-back).

Compare price strings only after normalizing — see "Site-Specific Gotchas" on price-string variance.

Site-Specific Gotchas

/obsessions.json does NOT exist — returns 404 with the TRR 404 page HTML. Verified iter 1. There is no JSON API surface for obsessions; HTML parsing is the only path. Don't waste turns probing /api/v1/obsessions, /api/obsessions, /users/me/obsessions, etc. — the only /api/* paths the site exposes are analytics collectors (/api/v1/collector/noScript.gif).
Anonymous /obsessions is NOT a 401/redirect — it returns 200 with an empty-state HTML page (heading "Obsessions", "Tap the ♡ next to any item…" copy, js-empty-obsessions-message, and a "Shop by Category" carousel). The HTML deliberately doesn't distinguish "logged out" from "logged in with no obsessions". The only auth signal is whether the page contains plp-product/ testids — if absent AND js-empty-obsessions-message present, the agent cannot tell the two states apart from HTML alone. If you need certainty, GET /account or /users/edit with the same cookies and check whether it 200s or redirects to /sign_in.
PerimeterX gates /sign_in with a PressAndHold human challenge for automated sessions (verified iter 1 — a stealth + residential-proxy session lands on https://www.therealreal.com/sign_in → "Access to this page has been denied" → PressAndHold iframe). Do not attempt scripted login. Inject the user's already-signed-in _session_id cookie via a setCookie command instead. Reference ID surfaced in PerimeterX response: d7e971a0-… — these are searchable in TRR's logs if the user reports lockout.
PerimeterX flags persistent sessions after the first navigation in many cases. In iter 3, a stealth + residential-proxy session loaded /shop/women/handbags cleanly, but the very next /obsessions request in the same session returned "Access to this page has been denied". So issue one browserless_agent call per page/request as a distinct session — because a call reusing the same session config reconnects to the same (flagged) session, don't reuse the prior config, so each call lands on a new residential IP (burn-and-rotate) — rather than reusing one long-lived session for a monitoring sweep. A single-navigation call worked first-try on iter 1 (200 OK with 30 KB of HTML and full PLP markup); reused sessions degraded after 1–2 navigations.
Three distinct price fields per card — pick the right one. price-final is the current asking price (what the user pays today). price-original is TRR's prior price for the same listing (struck-through, only present when discounted). price-msrp is "Est. Retail $X" — TRR's estimate of new-retail MSRP for the same item from the original brand; it does NOT change when TRR discounts. For "monitor pricing changes," anchor on price-final. The "Now N% off" callout is computed from price-final / price-original, not from MSRP.
Price strings include both $1,495 and $1,495.00 formats in the same response (compare the listing-grid card to the product detail card). Normalize before diffing: strip $ and ,, parse to float. Iter-1 sample on the handbags PLP: 482 $-prefixed strings, some without cents ($2,200), some with ($2,200.00). Don't assume a single canonical format.
Product IDs in data-testid are internal Rails IDs, not the user-facing slug. The slug (e.g. tumi-nylon-tote-u7x74) is in the -link href. Both are stable per-listing; use product_id as the diff key — slugs can change if TRR re-categorizes an item (e.g. moves from women/handbags/totes to women/handbags/shoulder-bags).
Image CDN URLs carry transform params. Each <img srcSet> is …?auto=webp&width=NNN&quality=40 NNNw, …. The canonical untranformed URL is https://product-images.therealreal.com/{SKU}_{N}_enlarged.jpg — strip everything after .jpg. The first part of the filename (TMI70173, LOU1250459) is the consigner SKU and is unique per physical item.
Items in the obsessions list can become "sold" without being removed. A delisted item silently disappears from the HTML grid; TRR does NOT render a "sold" tombstone in the obsessions PLP (the JSON-LD availability field on the detail page would still indicate OutOfStock, but that requires a per-item enrichment fetch). The cheapest "did this sell?" check is just: was product_id present last run and absent this run? Then enrich that one detail URL to confirm via availability.
The page is server-rendered Rails PJAX, not React. Don't wait for client-side hydration. An evaluate reading the DOM immediately after a goto with waitUntil: "load" returns complete data; no waitForTimeout step is needed before extraction.
The snapshot method collapses product cards into ARIA refs and drops data-testid attributes. Iter 1 confirmed: the anonymous empty /obsessions snapshot rendered the entire "Shop by Category" carousel as 6 distinct link refs but folded all the per-product data attributes out. For PLP extraction, always parse the DOM in-page with an evaluate selecting on data-testid. The snapshot method is fine for navigation refs (clicking specific buttons) but useless for enumerating tile data.
Don't tap the ♡ heart button — that's the un-obsess action and would remove items from the user's list. The data-testid="plp-product/{id}-obsession-button" element is read-only signal only; don't click it.
No skill verification against a real authenticated account was possible from this sandbox. All extraction details above were reverse-engineered from anonymous /obsessions HTML + the parallel /shop/women/handbags PLP markup (TRR reuses the same product-card component on both pages — verified by identical data-testid="plp-product/{id}-…" selectors in /tmp/skill/shop-fetch.json). When the agent first runs this skill with real user cookies, validate one extracted item against its product detail page's JSON-LD offers.price before trusting the rest of the batch.

Expected Output

{
  "captured_at": "2026-05-20T19:55:10Z",
  "total": 12,
  "page_count": 1,
  "auth_uncertain": false,
  "items": [
    {
      "product_id": "50758096",
      "designer": "Tumi",
      "name": "Nylon Tote",
      "url": "https://www.therealreal.com/products/women/handbags/totes/tumi-nylon-tote-u7x74",
      "image": "https://product-images.therealreal.com/TMI70173_1_enlarged.jpg",
      "price_final": "$116.00",
      "price_final_numeric": 116.0,
      "price_original": "$145",
      "price_original_numeric": 145.0,
      "price_msrp": "Est. Retail $480",
      "price_msrp_numeric": 480.0,
      "sale_callout": "Now 20% off",
      "obsession_count": "22"
    }
  ]
}

Empty-or-unauthenticated outcome

{
  "captured_at": "2026-05-20T19:55:10Z",
  "total": 0,
  "page_count": 0,
  "auth_uncertain": true,
  "items": [],
  "note": "HTML returned the empty-obsessions placeholder. Cannot distinguish 'logged in with zero obsessions' from 'session cookie invalid' from the HTML alone. Re-inject a fresh _session_id cookie and retry; if still empty, verify by hitting /account with the same cookie."
}

Diff outcome (when comparing to a prior snapshot)

{
  "captured_at": "2026-05-20T19:55:10Z",
  "prior_captured_at": "2026-05-19T19:55:10Z",
  "changes": [
    {
      "product_id": "50758096",
      "kind": "price_dropped",
      "from": 145.0,
      "to": 116.0,
      "callout": "Now 20% off"
    },
    {
      "product_id": "51049484",
      "kind": "new_callout",
      "callout": "Final Sale"
    },
    { "product_id": "51460206", "kind": "delisted", "last_seen_price": 1495.0 }
  ],
  "unchanged_count": 9
}

TheRealReal Obsessions Snapshot

Purpose

When to Use

Workflow

1. Acquire an authenticated session

2. Fetch `/obsessions`

3. Detect empty vs populated

4. Extract each item from `data-testid="plp-product/{id}-…"` markup

5. Paginate (if > 1 page of obsessions)

6. (Optional) Enrich per-item via JSON-LD

7. Diff against the prior snapshot

Site-Specific Gotchas

Expected Output

Empty-or-unauthenticated outcome

Diff outcome (when comparing to a prior snapshot)

Call it

Cookie Preferences

TheRealReal Obsessions Snapshot

Purpose

When to Use

Workflow

1. Acquire an authenticated session

2. Fetch /obsessions

3. Detect empty vs populated

4. Extract each item from data-testid="plp-product/{id}-…" markup

5. Paginate (if > 1 page of obsessions)

6. (Optional) Enrich per-item via JSON-LD

7. Diff against the prior snapshot

Site-Specific Gotchas

Expected Output

Empty-or-unauthenticated outcome

Diff outcome (when comparing to a prior snapshot)

Call it

Cookie Preferences

2. Fetch `/obsessions`

4. Extract each item from `data-testid="plp-product/{id}-…"` markup