[yummybomb]

Summary

• Adds five amazon.com intents that run the read-only tools from Add Amazon shopping and guarded retail cart tools #56 (search_products, get_product, get_best_sellers, get_cart). The bench registers only readOnlyHint: true tools, so Add Amazon shopping and guarded retail cart tools #56's cart-write tools stay out.
  • amazon-search-first-result (find): the first non-sponsored result for "usb c cable".
  • amazon-product-details (inspect): title, price, seller, rating and rating count for one ASIN.
  • amazon-best-sellers-top3 (find): the top three Kitchen & Dining Best Sellers, with an asin agreement check.
  • amazon-cheapest-well-rated (compare): the cheapest result rated 4.5 stars or more among the first 10 non-sponsored results, with a price lower-is-better agreement check.
  • amazon-return-window (gap): the standard return window, pinned to 30 days. No tool covers it, so it is the placebo.
• Adds the answer schemas these intents need.
• Changes the README pilot model to anthropic/claude-haiku-5-5 --thinking medium, and makes ANTHROPIC_API_KEY the example key.

Amazon shows a plain Kernel browser a "Continue shopping" interstitial, so run these intents with --stealth.

Runs

All three runs: --domain amazon.com --adoption 4 --pairs 5 --slots 4 --stealth, with #56 pinned at 12b5b74. Each run scored 54 trials and aborted none.

intent	gpt-5.4 speedup / adoption	haiku-5-5 medium	haiku-5-5 low
search-first-result	3.2× faster, 100%	0.7×, 60%	0.7× slower, 60%
product-details	3.3× (no signal), 100%	0.6×, 40%	0.8×, 20%
best-sellers-top3	3.2× faster, 100%	1.0×, Δ correct −2	0.8×, Δ correct −2
cheapest-well-rated	3.5× faster, 100%	0.8×, 80%	0.8×, 80%
return-window (gap)	0.6×, 0%	0.9×, 0%	0.8×, Δ correct −1
spend	$5.48	$0.24	$0.17

• gpt-5.4: every trial was correct in both arms. Treatment invoked a tool in every trial on the four tool-covered intents and was faster in 5 of 5 pairs on three of them.
• Haiku 5.5: control is already fast (median 13–60s per intent), so the extra turns spent listing and invoking tools make treatment no faster. It also skips the tools more often. Medium invoked a tool in more trials than low and got every gap-intent answer right, so the pilot uses medium.
• Best Sellers: Haiku's treatment misses are all on this intent. get_best_sellers returns price: null for some ranked items, and Haiku passed the null through instead of calling get_product, as gpt-5.4 did.

Testing

• node --test "bench/**/*.test.mjs": 97 pass.