agent0ai/space-agent/browser-control

Use this skill once at least one browser surface is open. browser-manager handles stand-alone window open or close work.

transient

• currently open web browsers lists browser id|url|title
• last interacted web browser may include fresh page content↓ with typed ref markers such as [link 12], [button 18], [image 24], or [input text 30]

workflow

• pick the target from transient or an explicit numeric browser id such as 1
• space.browser helpers already settle navigation state internally; do not call a separate sync step
• open, navigate, history, reload, state(...), content(...), detail(...), and dom(...) settle internally and return fresh read results without a separate sync step, and evaluate(...) still resolves a ready guest without requiring a separate sync step first
• ref-targeted action helpers such as click, type, submit, typeSubmit, and scroll return { action, state }; inspect result.action.status before retrying
• open(...) and typed navigate(...) treat bare hosts like novinky.cz or localhost:3000 as browser-address input instead of app-relative paths
• widget or page-authored <x-browser src="google.com"></x-browser> surfaces appear in the same browser list as stand-alone windows
• inspection helpers such as content(...), detail(...), dom(...), and state(...) also settle internally and mark that browser as the current prompt-time page-content source
• prefer the fresh last interacted web browser transient block after open, navigate, history, reload, and ref-targeted actions before asking for another explicit content(...) capture
• use content(...) to get readable page content with stable typed refs for the latest capture; unlike dom(...), it should stay cleaned up for agent use and should not include raw helper wrapper markup from nested frames or shadow roots
• dom(...) and content(...) accept either { selector: "..." } or { selectors: ["...", "..."] } when you want a scoped read instead of the whole page
• content(...) is lean by default to save tokens: it uses typed ref boxes like [link 12] Story, [disabled muted button 18] Continue, [checked checkbox 7] Email updates, or [input text 30] Search placeholder=Hledat value=Ethereum, omits link destinations, omits quotes around labels, and omits list bullets while keeping list indentation
• those refs also cover generic controls wired through framework or inline handlers such as @click, x-on:click, v-on:click, or onclick
• state and semantic tags are best-effort hints, not absolute truth; use detail(id, ref) when actionability is unclear
• if you need flatter output, content(id, { includeStateTags: false, includeSemanticTags: false }) suppresses those extra bracket tags
• images also get refs now, so detail(id, ref) works for image targets too
• if a specific link or image target matters, prefer detail(id, ref) on that reference instead of asking content(...) to print every -> url
• use detail(...) for deeper DOM on one ref before acting, and use it to inspect a link's real href, an image source, another referenced DOM target, or richer state metadata when the destination or actionability matters
• use click, type, submit, typeSubmit, and scroll only with refs from the latest content(...) capture
• typeSubmit(...) types into the field and then presses Enter in that same field
• any new content(...) call or navigation replaces the old ref ids
• prefer high-level helpers first; use evaluate(id, script) or send(id, "evaluate", { script }) only as a last resort when refs or navigation helpers cannot express the step, and remember that the last evaluated expression or resolved promise value becomes the result
• if result.action.status.noObservedEffect === true, stop retrying the same action and re-read the page or inspect the relevant control with detail(...)
• treat validationTextAdded, nearbyTextChanged, descriptorChanged, valueChanged, checkedChanged, selectedChanged, and semanticHints as evidence about what the page did after your action
• use numeric ids like 1, not browser-1

main helpers

• discovery: list(), ids(), count(), has(id), state(id)
• navigation: navigate(id, url), reload(id), back(id), forward(id)
• inspection: dom(id, payload?), content(id, payload?), detail(id, referenceId), evaluate(id, scriptOrPayload)
• interaction: click(id, ref), type(id, ref, value), submit(id, ref), typeSubmit(id, ref, value), scroll(id, ref)
• escape hatch: send(id, type, payload?)

runtime notes

• in the packaged native app, bridge-backed reads and ref actions work through the injected browser runtime
• in ordinary browser sessions, guarded calls return a structured warning object and also log the same warning text to the console instead of partial native-only behavior

examples Checking browser 1 now _____javascript return await space.browser.state(1)

Reading refs from browser 1 now _____javascript return await space.browser.content(1, { selectors: ["main", "article"] })

Reading one scoped region now _____javascript return await space.browser.dom(1, { selector: "[role='tab'], .nav-link, .tab, button" })

Inspecting one link target now _____javascript return await space.browser.detail(1, 79)

Opting into fuller link-heavy output now _____javascript return await space.browser.content(1, { includeLinkUrls: true, includeLabelQuotes: true, includeListMarkers: true })

Opening the first referenced result now _____javascript const content = await space.browser.content(1) console.log(content.document) return await space.browser.click(1, 79)

Typing into the active search box now _____javascript const content = await space.browser.content(1) console.log(content.document) return await space.browser.typeSubmit(1, 79, "Space Agent")

Running a last-resort page script now _____javascript return await space.browser.evaluate(1, const target = [...document.querySelectorAll("*")] .find((element) => element.textContent.trim() === "Update") if (!target) { "no update tab found" } else { target.click() "clicked: " + target.tagName } )