[yummybomb]

Summary

Pins the bench intents of the open tool PRs to their branches, so each PR's tools are what treatment registers, as the amazon.com, fandango.com and ipwho.is intents already do.

• aa.com → aa.com: add flight status, award fares and itinerary previews #46, amazon.in → Add Amazon India catalog and guarded guest cart tools #62, booking.com → booking.com: keep the party's children through rates, selection and checkout #129, delta.com → delta.com: add route schedules and advertised fare handoffs #67, doordash.com → Add DoorDash store and account-checked cart tools #57, ebay.com → ebay.com: search, listing and seller tools with safe failures #80, google.com → Add Google Finance, Maps planning and Flights exploration tools #61, opentable.com → Add OpenTable availability, account reads and reservation preparation #81, shopify.com → Add shopify.com App Store search_apps and get_app tools #59, target.com → target.com: add availability, reviews and fulfillment preparation #68, ticketmaster.com → Add Ticketmaster attraction lookup and public ticket-choice preparation #51, united.com → united.com: add flight status, fare reads and itinerary handoff #71, vividseats.com → Add Vivid Seats discovery and exact ticket preparation #52, walmart.com → Add guarded Walmart shopping reads and cart actions #63.
• Where the PR and main both changed a tool file since the PR branched (search_flights.js on aa.com, delta.com, google.com and united.com), main is listed after the PR so its copy wins and only the PR's new tools are added, which is what a rebase will have to do.
• tasks/doordash.com/domain.json from Add DoorDash store and account-checked cart tools #57: set_delivery_address as an action and the two account-cart reads as needs_sign_in, which doordash-pizza-near-address relies on since bench: grade the DoorDash address from the header, and add Shopify App Store and Steam comparison intents #136.
• fandango-nearest-theaters dates itself relative to the run, like fandango-theater-showtimes; its fixed date had passed and stopped any run that selected the domain. Verity (2026) is still on sale.
• The tasks test's "intents default to main" check now uses an intent that has no tools entry.

Runs

Haiku 5.5 and GPT-6 Luna, both --thinking low, --adoption 3 --pairs 5 --slots 5, 2026-10-26. Plain browsers for the 14 domains above minus the checkout intents; --stealth for amazon.com and the ebay, booking, target and walmart checkout intents; a second stealth pass for the four sites the plain run could not reach. 272 + 105 + 48 trials per model, 0 aborted, about $1.80 (Haiku) and $0.90 (Luna).

PR	Haiku 5.5 low	GPT-6 Luna low
#46 aa.com	5/5 vs 4/5, 7.1× faster	4/5 vs 0/5
#56 amazon.com (stealth)	5/5 on all five, 1.5–1.7× on the tool intents	5/5 on all, 1.5–6.3×
#57 doordash.com	5/5 both, 3.1×	5/5 vs 1/5, 1.9×
#58 fandango.com	5/5 both, 5.3× and 2.8×	5/5 vs 0/5; 5/5 vs 4/5, 6.4×
#59 shopify.com	compare 4.0×, listing 1.6×	compare 1.7×
#61 google.com	5/5 both, 3.1×	5/5 vs 0/5
#67 delta.com	5/5 vs 2/5, 3.4×; cart 5/5 vs 3/5	4/5 vs 0/5; 2/5 vs 0/5
#129 booking.com	five reads 5/5, 1.5–3.3×; checkout 5/5 vs 2/5	cheapest-room 5/5 vs 0/5; checkout 0/5 both
#81 opentable.com	plain 3/5 vs 5/5; stealth 4/5 vs 5/5, 5.1×	plain 5/5 vs 1/5; stealth 5/5 vs 1/5
#52 vividseats.com	1/5 vs 5/5	1/5 vs 1/5
#80 ebay.com (stealth)	5/5 vs 4/5, 0.7×	0/5 both
#71 united.com	3/5 vs 3/5 with main's search_flights	1/5 both
#62 amazon.in, #51 ticketmaster.com, #68 target.com, #63 walmart.com	blocked by the site in plain and stealth (see below)	same

• Bot protection: Target (PerimeterX press-and-hold) blocks search_products in plain and stealth; amazon.in answers its automated-access page to every tool call in both; Ticketmaster pauses every plain session and 403s about a third of stealth calls; OpenTable refuses the availability request in plain browsers while the page widget works, and clears in stealth; Walmart's "Robot or human?" stops both arms at the cart. The affected tools report these as challenge/blocked errors rather than empty results.
• Luna low answers empty JSON without tools on most multi-step tasks and stops early on checkout flows, so its control arms are mostly 0/5.
• vividseats: get_ticket_offers returns not_ready while the site's JSON search 404s early in a session, and its advice to stop after three tries made Haiku give up where control read the listings panel.

Testing

• node --test bench/test/tasks.test.mjs: 28 pass.
• The runs above; reports are in bench/results/<run_id> on the machine that ran them.

[yummybomb]

Closing: the pins were only needed for a one-off QA run.