feat(skills): add checkout-browser companion skill (browse → pay with Link)

[shubh24]

What

Adds a new companion skill, checkout-browser (skills/checkout-browser/SKILL.md), that gives an agent the ability to browse a real merchant to checkout and then pay with a Link one-time card.

Today create-payment-credential assumes the agent can already "navigate to the merchant page" and "enter card details into the checkout form" — but it doesn't provide that browsing capability. This skill is that missing half. The two compose cleanly:

• create-payment-credential → the payment flow (auth → spend-request → retrieve credential)
• checkout-browser → the browsing flow (navigate the site → scrape the live total → hand off to create-payment-credential → fill the card → place the order)

It's authored by Browserbase and driven by the browse CLI. Opening as a draft for your feedback on fit, placement, and conventions.

Scope of this PR

• One file added: skills/checkout-browser/SKILL.md. No source, README, CLAUDE.md, or manifest changes.
• It ships automatically via the existing plugin — plugins/link/skills is a symlink to skills/, so no registration change is needed.
• The skill is independently versioned (0.1.0) and is not wired into scripts/sync-skill-version.js. Happy to align it to the package version / add it to the sync script if you'd prefer skills track the package.

Design

• Recipe-first, own-flow fallback. Step 1 checks the Browse.sh catalog for a site recipe (browse skills find); if one exists it's a warm-start, otherwise the agent runs a generic search → product → cart → checkout flow.
• Local or remote. browse --remote (Browserbase cloud: stealth, CAPTCHA solving, proxies, recording) or --local/--auto-connect (the user's own Chrome). Guidance on when to pick each is in the skill.
• Total-match guardrail. The agent authorizes the scraped live total via spend-request, never an estimate, and stops if it exceeds the user's ceiling.
• Human-in-the-loop. The card only issues after the user approves the exact amount in the Link app; the agent never auto-places an order the user hasn't seen.

Validated end-to-end

This was tested live before opening, including a real purchase:

• Real Amazon checkout — searched, picked a well-rated item, drove cart → checkout, scraped the live total ($10.85), minted a Link one-time card for that exact amount (approved in the Link app), added it as the Amazon payment method, and placed the order. ✅
• Generic own-flow — full cart → checkout → order-review total on a sandbox store (no recipe), proving the fallback path.
• Recipe sweep — across 7 top merchants, recipes exist but are read-only search/extract; this is why the skill treats recipes as a warm-start, not a checkout shortcut.

A couple of real gotchas surfaced and are documented in the skill — most notably that cross-origin card iframes (Amazon's add-a-card form, Stripe Elements) block same-document fill, but the accessibility tree crosses frames, so browse snapshot + fill-by-@ref reaches the fields.

Open questions for maintainers

1. Placement — happy for this to live here, or in browserbase/browse.sh if you'd rather not host a third-party skill in-repo.
2. Versioning — keep it independently versioned, or sync to the package version like create-payment-credential?
3. Anything you'd want changed in tone/structure to match house conventions for skills.

I'll sign the CLA. 🤖 Generated with Claude Code

[cla-assistant]

All committers have signed the CLA.