[yummybomb]

Stacked on #118.

Summary

Bench intents now describe what a user wants done on a site, not which tools exist. The bench measures the speedup where tools exist, and shows where on the site agents spend their time, so tools can be built for the steps that need them.

Task-first intents

• tools is optional (main by default), and tools[].names and entry_tool are optional. Treatment gets every read-only tool the domain has at the pinned commits, plus the domain's actions (and checkout_actions) from bench: checkout intents graded from the page, with a submission guard #118 where the intent allows them. The start page is checked against the domain, or against the entry tool's patterns when an intent names one. Treatment setup checks that every tool registered.
• The checkout intents from bench: checkout intents graded from the page, with a submission guard #118 drop their tools and entry_tool.
• A task on a domain with no tools runs control alone: no adoption trials, no treatment, one trial per pair.

Where the time goes

• Each Kernel trial reads its top-frame navigations from browser telemetry and records agent time per step (page), with the calls made and WebMCP tools invoked on each. An intent's optional steps name pages by URL pattern; otherwise a step is host and path with id-like segments collapsed. delta-cart-main-classic names its search, results and Trip Summary steps.
• The report adds, per intent, a steps table (seconds per trial in each arm, how many treatment trials invoked a tool there, and which tools); a table for tasks with no tools yet; and a tool backlog of steps where control spent at least 20s per trial and treatment invoked a tool in fewer than half its trials, or that have no tools at all.
• kernel browsers telemetry events --all with --categories or --types stops after the first page with no matching event (it returned 8 of 24 navigations on a 6-minute trial), so the runner follows next_offset itself.

Limits

• Navigations within a single-page app (pushState) aren't recorded, so their time lumps into one step. Blue Bottle's storefront shows this.
• Local runs record no steps.

Testing

• node --test "bench/**/*.test.mjs": 97 pass on Node 26.
  • New tests: step names and the timeline, task-first validation, default sources, the coverage tables and backlog.
  • Run tests against stubs: a task naming no tools registers the domain's tools and records steps; a domain with no tools runs control alone, then has nothing to resume.
• Real runs on Kernel, pi with openai/gpt-5.4-mini:
  • delta-cart-main-classic: treatment correct in 86s. Control hit the turn cap after 409s on the results page; the steps table shows that, and that treatment used search_flights, list_flight_results, select_fare and select_fare_experience there.
  • bluebottle-checkout-espresso: treatment correct in 41s; control hit the turn cap.
• The page-based grading these runs used moved down into bench: checkout intents graded from the page, with a submission guard #118.