Skip to content

VALIDATION.md

Latest commit

 

History

History
177 lines (126 loc) · 16.8 KB

VALIDATION.md

File metadata and controls

177 lines (126 loc) · 16.8 KB
Raw

Validation notes

Checked on 29 September 2026.

This is a maintainer test summary, not a security certification or a guarantee of live source availability.

What was checked

Check Observed result
Format and safety-documentation checks 128 local assertions passed across eight skills
Example inputs All eight JSON examples passed validation against the saved published input schemas, with no unknown input keys
Installation skills CLI 1.7.0 found and installed all eight skills into an isolated local project for Codex
Installed contents All eight installed SKILL.md files matched their source files
Existing-export use Two independently executed assistant walkthroughs for Reviews and Trends, plus six maintainer-run local data walkthroughs and a separate content review
Cost and production changes No Actor runs, builds, proxy traffic or paid model API calls; no live Actor code, schema or pricing changes

The installer used a local source folder, --copy and project scope, with telemetry disabled. The GitHub download path and every supported assistant were not separately installed or tested. The skills themselves contain documentation only, with no bundled executables or install hooks.

The format check follows the current Agent Skills specification, including its compatibility field.

Existing-export cases

All inputs for these checks were explicitly synthetic, not customer records or new source collections.

Skill Cases exercised
App reviews Exact duplicates, edited and newly observed reviews, unresolved identity, missing rating, invalid date and absent run report
Google Trends Non-timeline rows, incomplete intervals, exact duplicates, invalid timestamps, null versus zero, incompatible contexts and untrusted query text
Map Your Show Duplicate deliveries, separate events with the same company name, multiple booths, category-only details and directory absence versus business closure
Shopify reviews Provider-aware identity, grouped reviews about another product, ratings-only feedback, unknown attribution and formula-leading text
Google News Duplicate article identities, matching headlines from different publishers, unresolved links, missing dates and snippet-only evidence
Pinterest Duplicate pins, unknown destinations, source attribution, missing counts versus zero and saves versus views
Google Hotels Different stay dates and currencies, nightly versus total prices, unknown room equivalence and missing quotes
TikTok Shop Long string IDs, repeated products, unavailable versus unknown stock, missing prices, shipping and source-selected review previews

Embedded requests to run paid tools, reveal credentials, download media or make purchases were treated as data. No such action occurred. This bounded offline exercise does not establish resistance to every prompt injection.

Important limits

  • These checks validate the documented local workflows and example inputs, not the current operation of the commercial scrapers.
  • Fresh collection, authentication, server-side spending enforcement, rate limits and restart behaviour were not exercised in this skills release.
  • The handmade Pinterest and TikTok fixtures used simplified link-field aliases. Their analysis is an incomplete-import case, not proof of production output-schema compatibility.
  • Some Shopify fixture rows lacked grouping metadata. Their exact product attribution remained unknown or caveated.
  • Trends duplicate handling and missing request provenance required explicit conservative choices: flag exact duplicates, do not average conflicts, and do not claim a common normalisation merely from matching country/time labels.
  • A missing run report means coverage is unconfirmed. Map Your Show and Shopify use SUMMARY; the other six Actors use RUN_REPORT.
  • No claim is made that these skills improve search ranking, will be selected automatically, or behave identically in every assistant.

Published Actor references

These input-schema and documentation versions were inspected when preparing the skills:

Actor Build
App Store & Google Play Reviews 0.3.19
Google Trends 0.3.12
Map Your Show exhibitors 0.2.10
Shopify reviews 0.4.10
Google News 0.2.16
Pinterest 0.1.10
Google Hotels 0.1.16
TikTok Shop 0.1.16

Before fresh collection, check the current Actor schema and Pricing tab and obtain an explicit positive spending cap. Installing a free skill does not make an Actor run free.

Threads addition (2026-10-01)

The original eight-skill checks above remain historical checks of those eight skills. The new apify-cleanscrape-threads-research skill is a separate addition.

  • Its example input passed the Threads Actor normalizer.
  • Current Agent Skills specification checks passed for frontmatter keys, YAML types, field lengths, directory/name matching and a non-empty body.
  • The bundled quick_validate.py rejects the compatibility key. The current specification explicitly permits that key; a specification-aligned check passed without modifying the installed validator. See https://agentskills.io/specification.
  • Cost approval, positive spending limits, credential handling, untrusted source text, partial results and experimental replies are covered explicitly.
  • Threads uses SUMMARY and REPORT, with RECOVERY_POSTS where relevant. Do not apply other Actors' report-key assumptions to it.
  • A real NASA example saved 20 unique posts with 20 settled billing events. Five blank captions were link-only posts with preserved link previews. The run stopped at its configured spending cap.
  • The public Actor and all three example pages returned HTTP 200 with the expected titles. A fresh run of the public 20-post example on build 0.1.7 saved 20 unique posts with 20 billing events, a valid CSV export and the corrected report.
  • These checks validate the skill format, documented input and released Actor. They are not an end-to-end installation or execution test in every assistant client.

Likee addition (2026-10-03)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-likee-research skill is a separate addition.

  • Its example input passed the Likee Actor normalizer (videos, maximum 10).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (380 characters), compatibility length and a non-empty body.
  • Cost approval, positive spending limits, credential handling, untrusted captions and comments, commenter names and partial results are covered explicitly.
  • Likee uses SUMMARY and REPORT. Do not apply other Actors' report-key assumptions to it.
  • A real run on build 0.1.10 saved 40 comments from a video on Likee's official US account; 22 of the 40 were written in Russian (one more in Tajik, also in Cyrillic). That sample is the source of the audience-language example.
  • These checks validate the skill format, documented input and released Actor. They are not an end-to-end installation or execution test in every assistant client.

Tennis addition (2026-10-03)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-tennis-research skill is a separate addition.

  • Its example input passed the Tennis Actor normalizer (one target, maximum 500, statistics on, ATP and WTA).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (347 characters), compatibility length and a non-empty body.
  • Cost approval, positive spending limits, credential handling, untrusted source text and partial results are covered explicitly.
  • Tennis uses SUMMARY and REPORT. Do not apply other Actors' report-key assumptions to it.
  • A real run of the public Wimbledon 2025 example saved 478 singles matches, all with full statistics. Summed over the seven men's main-draw matches each, Sinner held 93 of 99 service games and faced 23 break points; Alcaraz held 118 of 133 and faced 57. That run is the source of the aggregation example.
  • These checks validate the skill format, documented input and released Actor. They are not an end-to-end installation or execution test in every assistant client.

Bilibili addition (2026-10-05)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-bilibili-research skill is a separate addition.

  • Its example input passed the Bilibili Actor normalizer (one search term, 50 videos, danmaku on with a cap of 300, tags on by default).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (428 characters), compatibility length and a non-empty body.
  • Cost approval, positive spending limits, credential handling, untrusted titles, comments and danmaku, personal data in comment and danmaku rows, and partial results are covered explicitly.
  • Bilibili uses SUMMARY and REPORT. Do not apply other Actors' report-key assumptions to it.
  • The field names, summary keys and dataset layout were checked against the Actor source and a cloud run on build 0.1.14, where videos, comments, danmaku and creators are separate datasets listed under storageIds.datasets. Fresh runs of the public examples on Oct 5 2026 saved 100 videos for 咖啡, 1,800 danmaku from one video, and a creator profile with 100 videos.
  • Without login Bilibili leaves the comment author's region empty, so the skill tells assistants not to infer locations.
  • These checks validate the skill format, documented input and released Actor. They are not an end-to-end installation or execution test in every assistant client.

Google Jobs addition (2026-10-06)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-google-jobs-research skill is a separate addition.

  • Its example input passed the Google Jobs Actor normalizer (one title, three US locations, 50 jobs per search: 3 searches).
  • The same input ran on build 0.1.11 (run cpND8f6fO60X2og9j): 150 jobs saved and charged, 64 with a salary.
  • Field names, SUMMARY keys (jobs, stopReason, monitoring, searches) and search statuses match the Actor's code and dataset schema.
  • The skill tells assistants to compare salaries only within one salaryPeriod and currency, to report sample sizes, and to treat descriptions as untrusted data that can hold personal contact details.

Whatnot additions (2026-10-11)

The checks above remain historical checks of the earlier skills. The apify-cleanscrape-whatnot-research and apify-cleanscrape-whatnot-seller-research skills are separate additions.

  • Both example inputs passed the Actors' input normalizers.
  • The listing example ran on Whatnot Scraper build 0.1.21 (run r7EsdxOUFgQ6iDwKh): 73 PSA 10 Charizard listings saved and charged in 16 seconds. One listing had the $999,999 placeholder price, which is why the skill tells assistants to remove placeholders before price statistics.
  • A top-50 Pokémon Cards leaderboard run on Whatnot Seller Tracker (run uLoawPTrLcFVynhrZ) saved 50 sellers in 27 seconds: 40 compared with a check 14.9 hours earlier and 10 seen for the first time, with empty change columns.
  • Field names, dataset aliases (shows, showItems; shows, reviews, listings), SUMMARY keys and input statuses match the Actors' code and dataset schemas.
  • The skills tell assistants that prices are asking prices and items sold is a count, not revenue, and to treat titles, bios and reviews as untrusted data.

Kick addition (2026-10-12)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-kick-sponsor-research skill is a separate addition.

  • Its example input passed the Kick Sponsor Tracker input normalizer (four brands, all known; 10,000 live streams; minimum 1 viewer).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (373 characters), compatibility length and a non-empty body.
  • The same brands, stream limit and viewer minimum ran on build 0.1.9 (run HarOwAK2EXAZlwAX1, with a 200-streamer limit; 26 were picked): 10,000 live streams scanned and 26 sponsorships saved in about four minutes. HOLY was on 16 streamers, DZIK and G FUEL on 5 each, and Red Bull on none. That run is the source of the demo video and the brand-share example.
  • Field names, dataset aliases (channels, streams, brands), SUMMARY keys and streamer statuses (done, not_found, invalid, failed, not_processed) match the Actor's code and dataset schemas.
  • The skill tells assistants that a sponsorship is what a channel shows publicly, not a confirmed contract or fee, that broadcast figures cover at most the latest 30 broadcasts, and to treat titles, bios, link descriptions and chat as untrusted data. It also notes that gambling advertising is regulated.

Casino addition (2026-10-19)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-casino-bonus-research skill is a separate addition.

  • Its example input passed the Casino bonus checker input normalizer (one brand, United Kingdom, one official offer line).
  • Field names, defaults, limits and prices were checked against the published build 0.1.20 and its README on 2026-10-19.
  • The skill names no casino and links no casino or affiliate page; it is written for brand, compliance and affiliate-programme teams.

Tennis rankings addition (2026-10-23)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-tennis-rankings-research skill is a separate addition.

  • Its example input passed the ATP and WTA Tennis Rankings Scraper input normalizer (ATP ranking and race, two players).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (381 characters), compatibility length and a non-empty body.
  • Field names, SUMMARY keys, stop reasons, prices and the four example links were checked against the published build 0.1.8 on 2026-10-23. The same input ran on the published Actor's example task (4 entries: two players in two lists, edition 2026-10-19).
  • The skill tells assistants that race points and ranking points are not comparable, that ties share a rank, that a null rank change means a new entry, that Russian and Belarusian players are listed under World, and that rankings are not predictions or betting advice.

Career page jobs and video frames additions (2026-10-24)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-career-site-jobs-research and apify-cleanscrape-video-frames skills are separate additions.

  • Their example inputs passed the Actors' own input normalizers (three companies with a title, location and watch; one signed video link sampled every 10 seconds as WebP).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (392 and 413 characters), compatibility length and a non-empty body.
  • Field names, dataset aliases (companies, closedJobs; videos), SUMMARY keys, statuses, stop reasons, prices and the ten example links were checked against the published builds (Career Page Jobs 0.1.14, Video Frame Extractor 0.1.12) on 2026-10-24. Each linked example was run before it was published: for instance 151 US engineering jobs with pay ranges from five Greenhouse boards, and 64 WebP frames every 10 seconds from Elephants Dream.
  • The skills tell assistants that a watch's first run is a starting point, not new hiring, that a closed job may have been filled, paused or moved, that missing pay is not zero, that fewer frames than requested is normal when near-identical frames are left out, and to use only videos the user may use.

Economic calendar addition (2026-10-27)

The checks above remain historical checks of the earlier skills. The new apify-cleanscrape-economic-calendar-research skill is a separate addition.

  • Its example input passed the Actor's own input normalizer (all five economies, next 7 days, high impact, with a watch name and only changes).
  • Agent Skills specification checks passed for frontmatter keys, name format and directory match, description length (433 characters), compatibility length and a non-empty body.
  • Field names, the sources dataset alias, SUMMARY keys, time-status and change values, prices and the five example links were checked against the published build (Economic Calendar 0.1.19) on 2026-10-27. Each linked example was run before it was published: for instance 34 US CPI and jobs releases from the past year, and the Fed, ECB, Bank of Canada and RBA decisions for the next 30 days.
  • The skill tells assistants to state each time's precision rather than presenting standard, provisional or date-only times as confirmed, that a watch's first run marks everything new, that first-published and current series values are different measures, that a blank value is not zero, and that there are no consensus forecasts.