<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
parad0x-labs/vool-windows-m6a/vool-browser-qa
──Details
PublishedOctober 22, 2026 at 11:24 PM
Content Hashsha256:bc00cf28a11a6b11...
Git SHA78f818b802c5
──Files
Files (1 file, 3.3 KB)
SKILL.md3.3 KBactive
SKILL.md · 77 lines · 3.3 KB

name: vool-browser-qa description: "QA a web page or running site read-only: fetch the rendered page, verify expected content and structure, and report pass/fail per check. Use when the user asks to check a page, verify a site renders, or QA the web surface." version: 1.0.0 license: MIT author: sls_0x status: p1 type: capability_skill triggers: [browser, rendered, page check, verify the page, site qa, rendered page] allowed-tools: [web.fetch, web.search] capabilities: [web.read] permissions: [read_files] effects: [read_only] inputs: [url, expected_content list] outputs: [qa_checklist_results, fetched_page_evidence] stop-conditions: [every expected check has a pass or fail with fetched evidence, fetch refused or blocked - report the refusal, not a guess about the page] recovery: [if the rendered fetch fails, fall back to the raw fetch and label the evidence weaker, if a check needs interactivity this lane cannot do, report it as untested rather than inferring] cumulative-test-law: [tests/native_skills must pass before this package changes, selection phrase "the rendered page in the browser lane" must keep selecting this package, claims about a page must cite fetched evidence, never memory] id: browser-qa risk-class: read_only task-families: [integration_orchestration] capability-families: [web] tool-intents: [web.fetch, web.search] permitted-tools: [web.fetch, web.search] prerequisites: [] expected-outputs: [qa_checklist_results, fetched_page_evidence] verification: [browser_verified, evidence_cited] stopping-conditions: ["stop when every checklist row carries observed evidence — a row with no evidence is UNVERIFIED, never green"] incompatible-with: [] priority: 30


Browser QA (read-only lane)

You verify pages against expected content with FETCHED evidence. Every claim cites what the fetch actually returned.

Honest limits (read these first)

This package declares web.read only: fetched-page QA. INTERACTIVE browser driving (click, type, screenshot a live session) is NOT wired into the runtime at this version — the browser-render contract exists but has no executor lane behind it, and this package will not pretend otherwise. Interactive checks are reported as UNTESTED, never inferred.

Required capabilities (declaring is not granting)

web.read. Declaring is not permission: fetches go through the transport door with its size caps and refusal handling.

Procedure

  1. Collect the expected-content checklist from the user's ask.
  2. web.fetch the URL; keep the evidence (status, extracted text).
  3. Check each expectation against the FETCHED text — pass/fail per item.
  4. web.search only to locate the right URL when the user gave a name, not a

substitute for fetching the page itself.

  1. Deliver the checklist results with the fetch receipt.

Stop conditions

Stop when every checklist item has a verdict with evidence — including honest UNTESTED rows for interactive checks.

Recovery

Rendered fetch fails: fall back to raw fetch, label the evidence weaker. Fetch refused: report the refusal verbatim; never guess the page's content.

Cumulative test law

tests/native_skills is the pack; QA-behaviour edits re-run it.

No direct execution bypass

Page fetching goes through the runtime tool door (execute_runtime_tool) and its transport lane. No direct network reads around it.

All versions