missav

name	missav
description	Browse missav.ws AV listings with agent-browser, extract title/duration/url from listing cards, enrich detail-page plot descriptions for genre analysis, and rank recommendations against user preferences stored in ~/.missav/ (genres/types, performers, labels/studios, keywords, duration). Learns from user feedback each session. Use when finding AV on missav, searching /cn/search/{query}, analyzing plot descriptions, or getting personalized recommendations.
metadata	{"requires":{"bins":["agent-browser"]}}

missav

Browse missav.ws with agent-browser, extract listings, enrich detail-page descriptions, rank against ~/.missav/ preferences.

Prerequisite: ../agent-browser/SKILL.md, references/browser-troubleshooting.md

Install: npm i -g agent-browser && agent-browser install

Authorized access only: If the site presents a Cloudflare challenge, stop the workflow immediately. Do not retry challenges, alter browser fingerprints, rotate proxies, or use challenge-solving services. Use a site-authorized API or access path instead.

Preference store (~/.missav/)

node skills/missav/scripts/preferences.mjs init
node skills/missav/scripts/preferences.mjs show

See references/preferences-schema.md.

Dimension	Field	Example
类型	likes.genres	SM, NTR
男优/女优	likes.performers	从标题或详情页提取
厂牌	likes.labels	番号前缀 SSIS-825 → SSIS
关键词	likes.keywords	english-subtitle

Workflow (optimized)

Use run-recommend.sh for the full tested pipeline, or run steps manually.

The workflow is fail-closed: a Cloudflare challenge detected by title or body returns exit code 42, stops all later searches/detail requests, and never presents a partial recommendation as a successful run. Public search results and community-maintained clients are not a site-authorized API; do not replace this behavior with fingerprint changes, proxy rotation, challenge solving, or unofficial bypasses.

chmod +x skills/missav/scripts/run-recommend.sh
./skills/missav/scripts/run-recommend.sh --prefs ~/.missav/preferences.json

This script: loads prefs → searches each liked genre + label → merges listings → ranks → enriches top scored detail pages → re-ranks with descriptions.

Manual steps

Task Progress:
- [ ] Step 0: preferences.mjs init && show
- [ ] Step 1: missav_browser_start (once per task)
- [ ] Step 2: browse-page.sh URL --continue (for 2nd+ URLs)
- [ ] Step 3: extract-listings.sh → normalize-eval-output (built-in)
- [ ] Step 4: filter-recommendations.mjs → ranked.json
- [ ] Step 5: enrich-listings.sh --ranked ranked.json (top scored only)
- [ ] Step 6: filter again + present + feedback
- [ ] Step 7: missav_browser_close

Browser rules (from testing)

1. One missav_browser_start per task — then --continue for more URLs
2. Always use the bundled browser defaults in browser-env.sh
3. Never use wait --load networkidle
4. Treat a Cloudflare challenge as terminal for the current workflow; do not issue another search or detail request
5. Validate URL after open — reject /pop, /search/ redirects on detail pages
6. Pass --ranked to enrich, not raw listing order
7. Treat detail enrichment challenges as fatal; list-only fallback is reserved for ordinary missing-description/landing failures

Step 0–1: Preferences + URLs

node skills/missav/scripts/preferences.mjs show

# Build one URL per liked label/genre (merge results later)
node skills/missav/scripts/build-search-url.mjs search "{label}"
node skills/missav/scripts/build-search-url.mjs search "{genre}"

Search pattern: https://missav.ws/cn/search/{keyword}

Step 2: Browse + extract (use bundled browser defaults)

SESSION="missav-$(date +%s)"
export MISSAV_SESSION="$SESSION"
source skills/missav/scripts/lib/browser-env.sh
missav_browser_start

./skills/missav/scripts/browse-page.sh "https://missav.ws/cn/search/sm"
./skills/missav/scripts/extract-listings.sh --session "$SESSION" --output /tmp/listings-a.json

# 2nd query: reuse session, do NOT restart browser
./skills/missav/scripts/browse-page.sh "https://missav.ws/cn/search/omhd" --session "$SESSION" --continue
./skills/missav/scripts/extract-listings.sh --session "$SESSION" --output /tmp/listings-b.json || true

browse-page.sh handles: close stale daemon → open page → challenge detection → age gate → scroll. A detected challenge stops the workflow.

If extraction returns 0 listings, read browser-troubleshooting.md and retry once.

Step 3: Preliminary rank

node skills/missav/scripts/filter-recommendations.mjs \
  /tmp/missav-listings.json \
  --prefs ~/.missav/preferences.json \
  --limit 15 > /tmp/missav-prelim.json

Step 4: Detail enrichment (description analysis)

For top candidates, open detail pages and extract plot description:

<div class="mb-1 text-secondary break-all line-clamp-none">
  …剧情描述文本…
</div>

node skills/missav/scripts/filter-recommendations.mjs \
  /tmp/missav-merged.json \
  --prefs ~/.missav/preferences.json \
  --limit 10 > /tmp/missav-ranked.json

Step 4: Detail enrichment (top scored items only)

./skills/missav/scripts/enrich-listings.sh \
  --input /tmp/missav-merged.json \
  --ranked /tmp/missav-ranked.json \
  --output /tmp/missav-enriched.json \
  --top 3 \
  --session "$SESSION"

Single detail page:

./skills/missav/scripts/browse-page.sh "https://missav.ws/dm1/cn/{slug}"
./skills/missav/scripts/extract-detail.sh --session "$SESSION" --output /tmp/detail.json
node skills/missav/scripts/analyze-description.mjs \
  --item /tmp/detail.json \
  --prefs ~/.missav/preferences.json

Description analysis matches user likes.genres (e.g. SM) against plot keywords: 绑、深喉、受虐、调教…

Step 5: Final rank

node skills/missav/scripts/filter-recommendations.mjs \
  /tmp/missav-enriched.json \
  --prefs ~/.missav/preferences.json \
  --limit 10

Items with description get extra score via description ↔ preference matching.

Step 6: Present + feedback

## missav 推荐

| # | 番号 | 时长 | 匹配 | 标题 | 链接 |
|---|------|------|------|------|------|
| 1 | … | … | SM, 厂牌 | … | [打开](…) |

### 剧情摘要（Top 1）
{first 2 sentences of description}

### 偏好匹配
- 类型 SM：描述含「绑」「深喉」等
- 厂牌：{label}-xxx

On user feedback, persist with full item including description:

node skills/missav/scripts/preferences.mjs feedback like \
  --item /tmp/detail.json \
  --note "用户喜欢这类SM剧情"

feedback like auto-infers genres from description text into likes.genres.

Cleanup

source skills/missav/scripts/lib/browser-env.sh
missav_browser_close

Listing card DOM

<a href="…" alt="ssis-825-english-subtitle">
  <span class="absolute bottom-1 right-1 …">2:03:03</span>
</a>
<a class="text-secondary group-hover:text-primary" href="…" alt="…">
  SSIS-825 … Fuua Kaede - 枫ふうあ
</a>

See references/dom-selectors.md.

Error handling

Symptom	Action
Just a moment... title or Cloudflare challenge body	Stop immediately with exit code 42. Automated access requires a site-authorized API or access path.
Executable doesn't exist	Set MISSAV_CHROME to system Chrome
daemon already running	agent-browser close first
Empty [] extraction	Run browse-page.sh; check age gate + scroll
Empty description	Click 展开/更多; use extract-detail.js
JSON parse error	Use parse-eval-json.mjs (double-encoded fix)

Scripts

Script	Purpose
run-recommend.sh	One-shot full pipeline
browse-page.sh	Open URL (--continue for 2nd+ page)
normalize-eval-output.mjs	Fix double-encoded eval JSON
extract-listings.sh	Listing cards → JSON array
extract-detail.sh	Detail page → description JSON
enrich-listings.sh	Batch detail fetch (--ranked required)
analyze-description.mjs	Description ↔ prefs analysis
filter-recommendations.mjs	Score + rank
preferences.mjs	~/.missav/ CRUD + feedback
build-search-url.mjs	URL builder

Browser-side eval: extract-listings.js, extract-detail.js

Node .mjs scripts: zero npm dependencies.

References: preferences-schema.md, browser-troubleshooting.md