Ebay Item Detail

Known Limitations

• eBay regional geo-redirection: some regional TLDs (ebay.de, ebay.co.uk) will render the same item with localized subtitle wording ("Wird in neuem Fenster oder Tab geöffnet", etc.). The extractor strips these across the observed languages (English, German, French, Spanish, Italian, Dutch, Simplified Chinese, Japanese); other locales may leave residual accessibility suffixes in the raw text fields.
• Ended, sold-out, or removed listings often 302-redirect to a similar-item search page. The extractor detects this via URL mismatch and returns an error rather than scraping the wrong page.
• eBay obfuscates some class names on the sold-count line (ONPa, h1WV, etc. change over deploys). The extractor uses a text-pattern scan (^{n} sold$) instead of relying on those classes, which is stable across the observed language variants.
• Auction-specific fields (current bid, bid count, time-left) are not extracted by this component — they are part of the search-listing capability's per-card attribute row and are not surfaced consistently on the detail page.
• whyToBuy is a composed signal derived from other extracted fields (sold, wasPrice vs price, free-shipping flag, availability = last-one, top-rated seller heuristic). It is a convenience array; the underlying atomic fields are always present when meaningful.
• Extraction is DOM-based and reads only the initial server-rendered state. Fields that load lazily via later XHR (e.g., some seller reviews, live shipping estimates for the current buyer's ZIP) are not fetched — the extractor takes whatever the page shows at the moment of the eval.

Execution Efficiency

• Batch orchestration: Feed item URLs (e.g., from a listing extractor) into a serial loop; navigate → wait stable → eval. Do not parallelize within one browser session — eBay throttles rapid item-detail hits from the same fingerprint. Insert a small pause (1–3s) between items to keep the traffic pattern natural. To increase throughput, run multiple stealth browser sessions in parallel — each has an independent fingerprint so rate limits apply per session.
• Test before batch execution: After wiring the loop, first process 2–3 items and inspect the output to confirm currency, seller, and Item specifics parse correctly for the current target vertical; only then run the full batch.
• Reduce redundant pre-operations: Reuse the same session across many item URLs; do not re-open the browser per item. Cache resolved detail JSON keyed by itemNumber so a re-run resumes from the next uncollected item.
• Error resumption: Persist per-item results (tmp/items/{itemNumber}.json) as they arrive; on failure, restart the loop with the "already collected" set filtered out.

Experience Notes

Path: {working-directory}/browser-act-skill-forge-memories/ebay-scraper-ebay-item-detail.memory.md (working directory is determined by the Agent running the Skill, typically the project root or current working directory)

Before execution: If the file exists, read it first — it records unexpected situations encountered during past executions (e.g., a strategy has become ineffective); adjust strategy order accordingly.

After execution: If an unexpected situation is encountered (strategy became ineffective, page redesigned, anti-scraping upgraded, better path discovered), append a line: {YYYY-MM-DD}: {what happened} → {conclusion}

Normal execution does not write to the file. Do not record what item numbers were processed or how many fields were populated — those are task outputs, not experience.