review-top-addons

name	review-top-addons
description	Download the latest published anki-addons dataset from HuggingFace, take the top N highest-rated addons, cross-check every field (AnkiWeb links, GitHub, Anki Forum) against the live sources, and write a markdown report of discrepancies. Focuses especially on whether the GitHub repo and Anki Forum URL were correctly extracted from the addon page. Use when the user says "review top addons", "/review-top-addons", or asks to verify the top-rated addons' data for correctness.
allowed-tools	Bash(python3 *) Bash(gh api *) Bash(curl *) Bash(grep *) Bash(pip install *)
disable-model-invocation	true

review-top-addons

Pull the latest published Ya-Alex/anki-addons dataset, take the top N highest-rated addons, verify each addon's fields against the live sources, and write a markdown report of the real discrepancies. The single most important check is whether the GitHub repo and Anki Forum URL were extracted correctly from the addon page — that is the most fragile part of the pipeline, and everything in the github/forum blocks is derived from it. Work through the steps in order; report clearly at the end.

N is the first argument the user gives (e.g. /review-top-addons 20). If none is given, default to 10.

1. Fetch the dataset and select the top N addons

python3 .claude/skills/review-top-addons/select_top_addons.py <N>   # N defaults to 10

This downloads latest/parquet/data.parquet (the canonical default config on HuggingFace) and latest/metadata.json, sorts by anki_web.rating descending, and writes to ~/anki-addons-dataset/reviews/:

• top-<N>-selected.json — the selected records. Each record has id, anki_web{title, addon_page_url, rating, update_date, anki_version, branches, links, likes, dislikes, ...}, github (nullable), forum (nullable), plus two derived helper fields the script pre-extracts from the addon page HTML: _page_github_urls and _page_forum_urls (every github.com / forums.ankiweb.net URL actually present on the page).
• pages/<id>.html — the exact rendered page HTML the parser saw (the extraction ground truth).

Note the printed snapshot date — the dataset is a point-in-time snapshot, so counts will have drifted since then (see step 2). If the script fails because huggingface_hub/pandas/pyarrow are missing, pip install huggingface_hub pandas pyarrow and retry (or run ./pip_update.sh). Read top-<N>-selected.json to drive the rest of the review.

2. Cross-check each addon against the live sources

For each selected addon, verify the fields below and record any mismatch. Do the link-extraction checks first — they are the priority.

2a. Link extraction (highest priority)

The addon page is client-side rendered, so it cannot be fetched with a plain GET — use the captured page instead. Ground truth = the page's own links, already extracted for you as _page_github_urls / _page_forum_urls (and the full HTML at pages/<id>.html, greppable with grep -oE 'https?://github\.com/[^\"'\''<> )]+' ~/anki-addons-dataset/reviews/pages/<id>.html | sort -u).

• GitHub repo choice — when github is non-null, confirm the chosen github.user/github.repo really is the addon's own repository and appears in _page_github_urls. Pages often link several repos (the author's other projects, contributors' forks, unrelated repos in prose); flag if the parser picked the wrong one, picked a fork/third-party repo over the author's, or invented a user/repo that appears nowhere on the page.
• Missing GitHub — when github is null but _page_github_urls clearly contains the addon's own repo, flag it as a missed extraction.
• Forum URL — when forum is non-null, confirm forum.anki_forum_url corresponds to a URL in _page_forum_urls and that topic_slug/topic_id match that URL. Pages may link multiple forum topics/categories; flag if the wrong topic was chosen, or if forum is null while an obvious official support thread is present in _page_forum_urls.

2b. GitHub enrichment (only if github is non-null)

Confirm the repo identity from 2a is right, then verify the enrichment values:

gh api repos/<user>/<repo> --jq '{full_name, stargazers_count, pushed_at, archived}'
gh api repos/<user>/<repo>/languages

• A 404 means the extracted repo URL is wrong — a genuine, high-value bug.
• Check languages matches github.languages, and github.last_commit is not after the repo's real last push (a last_commit in the future vs. reality is an error; the live repo being newer than the snapshot is just drift).
• stars, action_count, tests_count: sanity-check magnitude only (stars drift over time — see below).

2c. Forum enrichment (only if forum is non-null)

curl -s "https://forums.ankiweb.net/t/<topic_slug>/<topic_id>.json" \
  | python3 -c "import sys,json; d=json.load(sys.stdin); print({'id':d.get('id'),'slug':d.get('slug'),'posts_count':d.get('posts_count'),'last_posted_at':d.get('last_posted_at')})"

• A 404 / error means the extracted forum URL is wrong — a genuine bug.
• Confirm the returned id/slug match forum.topic_id/forum.topic_slug. Treat posts_count/last_posted_at as drift-prone (see below).

Snapshot drift vs. real errors

The dataset is a snapshot as of the printed snapshot date. Values that naturally grow/advance over time — rating, likes, dislikes, stars, posts_count, last_commit, last_posted_at — will legitimately be higher/later live than in the dataset. Only flag these when they moved impossibly (live value lower than the snapshot by more than a rounding wobble, or a snapshot timestamp in the future relative to the live source). Always treat as genuine issues: wrong/misattributed GitHub repo or forum URL, 404 repo/topic, wrong/empty title, wrong languages, mismatched anki_version, malformed fields, or a github/forum block that is present but nonsensical.

Live re-verification of AnkiWeb scalar fields (rating, likes, update_date) is limited: the page is JS-rendered and can't be fetched directly, so verify those against the captured pages/<id>.html rather than a live GET, and note in the report that live AnkiWeb re-fetch was not performed.

3. Write the report

Write a markdown report to ~/anki-addons-dataset/reviews/top-<N>-review-<YYYY-MM-DD>.md (use today's date):

• Header — N reviewed, snapshot date, review date, dataset file (latest/parquet/data.parquet).
• Issues found — a top-level summary listing only the real discrepancies, link-extraction problems (wrong/missing/misattributed GitHub or forum URL) first and most prominently, then other genuine errors. If there are none, say so explicitly.
• Per-addon detail — one section per addon (id, title, addon_page_url) with a compact table of the checked fields: dataset value vs. live/page value vs. verdict (✓ / ✗ / drift). Note where a live check couldn't be performed.

Then print a concise summary in chat: how many addons were reviewed, the count of real issues by type, and the full path to the report file. Lead with the link-extraction findings.

Done

Report the outcome: number of addons reviewed, snapshot date, the list of genuine issues found (link-extraction problems first), and the path to the written report — or state clearly that all top-N addons checked out with no real discrepancies.