Naver Blog Browser

Reusable headless browser helpers for Naver search and blog pages. Other Korean stock skills (kr-naver-blogger, kr-naver-insight) depend on this skill for navigation and text extraction. Treat this skill as the single place that knows which Naver URLs actually render text in a headless browser.

When To Use

• Another skill needs to search search.naver.com for blog posts about a company or ticker.
• Another skill needs the recent post list for a Naver blogger.
• Another skill needs the full text of a single Naver blog post.
• The user is asking you to read a Naver blog page and summarize it.

Do not use this skill for general web browsing. Use gstack /browse for that. This wrapper exists only because Naver's blog routes are JS-heavy and need a specific URL shape to return text at all.

URL Patterns That Work

Patterns that do not reliably render text in headless mode:

• https://blog.naver.com/<id> returns an empty shell
• https://blog.naver.com/<id>/<logNo> redirects through JS and text extraction often returns empty. Use browse-naver.js's normalizeNaverUrl() to rewrite it into PostView.naver?blogId=...&logNo=... before visiting.

How It Works

scripts/browse-naver.js is a Node module (no external deps) that resolves a gstack browse binary in this order:

1. $GSTACK_BROWSE_BIN environment variable, if set
2. A skill-local vendored binary at vendor/gstack/browse/dist/browse
3. (Mac/Linux only) The binary installed under the Codex skill location: $CODEX_HOME/skills/kr-naver-browse/vendor/gstack/browse/dist/browse (default ~/.codex/skills/kr-naver-browse/vendor/gstack/browse/dist/browse)
4. (Mac/Linux only) A gstack Codex skill installation: $CODEX_HOME/skills/gstack/browse/dist/browse
5. (Mac/Linux only) The binary installed under the Claude Code skill location: $CLAUDE_HOME/skills/kr-naver-browse/vendor/gstack/browse/dist/browse (default ~/.claude/skills/kr-naver-browse/vendor/gstack/browse/dist/browse)
6. (Mac/Linux only) A gstack Claude skill installation: $CLAUDE_HOME/skills/gstack/browse/dist/browse

Steps 3–6 allow running scripts directly from the source repo when the skill has been installed into Codex or Claude Code but the repo's own vendor/ directory has not been built. Each helper spawns the resolved binary via execFileSync with a 30s timeout and a 5 MB output buffer. Korean search terms are passed through encodeURIComponent. On empty text extraction, helpers retry once after a 5 s wait before returning null.

kr-naver-browse ships a scripts/post-install.(sh|ps1) hook. When the skill is installed through this repo's scripts/install-*.{sh,ps1} helpers, the hook builds or refreshes gstack under the installed skill's vendor/gstack/, ensures bun can produce runnable compiled executables (trying Homebrew, npm, then the official Bun installer pinned by BUN_VERSION_TAG, default bun-v1.3.10), runs bun install, bun run build, Linux bunx playwright install-deps chromium unless SKILL_INSTALL_SKIP_LINUX_DEPS=1, and bunx playwright install chromium, strips vendored SKILL.md files so Codex does not recursively load gstack's own skill pack, and verifies that vendor/gstack/browse/dist/browse can launch and navigate to https://example.com.

Install paths:

• Default Codex install: bash ./scripts/install-all-skills.sh
• macOS Codex recovery path: bash ./scripts/install-codex-mac-naver.sh

On macOS Codex installs under $CODEX_HOME/skills/ (default ~/.codex/skills/), if the compiled binary build finishes but the smoke test still fails, the hook automatically falls back to a Bun source wrapper at the same final path: vendor/gstack/browse/dist/browse. The higher-level Naver skills keep calling the same path either way.

Set SKILL_INSTALL_SKIP_HOOKS=1 to skip this bootstrap, set SKILL_INSTALL_SKIP_LINUX_DEPS=1 to skip automatic Linux/WSL system dependency installation, set SKILL_INSTALL_AUTO_BUN=0 to avoid automatic bun installation, set SKILL_INSTALL_FORCE_CODEX_MAC=1 to force the macOS Codex fallback path during install, or set GSTACK_BROWSE_BIN to an absolute existing gstack browse binary path.

Helpers Exposed

const {
  resolveBrowseBinary,
  browseText,
  searchNaverBlog,
  readBlogPostList,
  readBlogPost,
  normalizeNaverUrl,
  parseNaverSearchResults,
  fetchRelatedQueries,
  extractBlogIdRefs,
} = require("../../kr-naver-browse/scripts/browse-naver.js");

• resolveBrowseBinary() returns the absolute path to the gstack browse binary, or throws with a clear install hint.
• browseText(url, { timeoutMs }) visits the URL, waits for the page to settle, and returns the extracted text or null on failure.
• searchNaverBlog(query, { max }) returns an array of search result objects { blogId, title, snippet, date, url }.
• readBlogPostList(blogId) returns recent posts for a blogger as an array of { logNo, title, date }.
• readBlogPost(blogId, logNo) returns { title, date, text, url } or null.
• normalizeNaverUrl(url) rewrites blog.naver.com/<id>/<logNo> variants into the PostView.naver shape.
• parseNaverSearchResults(text) is a pure parser over raw search text, useful for testing.
• fetchRelatedQueries(company) scrapes the 연관검색어 block from the blog search page for company and returns [{ text, nickname, wholeQuery }] — the authoritative Naver-side signal for "who else is searched alongside this ticker". Used by kr-naver-blogger to surface famous bloggers by nickname. Results are deduped and capped at 20.
• extractBlogIdRefs(pageText, { excludeBlogId }) scans a page text dump for blog.naver.com/<blogId> URLs and returns a deduped list of blogIds, filtered against a small blocklist of reserved path segments (PostView, PostList, api, ...) and optionally excluding a caller-supplied blogId (the source post's own author). Used to mine roundup-post bodies for referenced bloggers.

Operating Rules

1. Always normalize URLs before calling browseText. A raw blog.naver.com/<id>/<logNo> URL will silently return empty text.
2. Sleep 1 second between requests to the same Naver host. The helpers do this automatically; if you call browseText directly in a loop, enforce the delay yourself.
3. Retry once on empty extraction, then give up. Do not hammer the page. Naver is quick to rate-limit.
4. Never scrape personal info (subscriber lists, contact details). Only public post content and public metadata (title, date, logNo).
5. Truncate text to 3000 chars before returning post bodies up the stack. Callers never need the full post for summarization and the truncation keeps downstream LLM context small.
6. Fail loud, not silent. If the browse binary is missing or invalid, throw with the install hint or rerun the skill installer so the post-install hook can bootstrap a local binary. Do not fall back to fetch; dynamic pages will not render.

Smoke Test

node skills/kr-naver-browse/scripts/browse-naver.js --test

The --test flag runs a single search for 엘앤에프 주가, prints the first three results, and exits non-zero if the browse binary is missing or returns empty text both times.

Known Issue

• Sandbox runtime bug: some Codex sandbox environments block local 127.0.0.1 listen calls with EPERM. When that happens, the underlying gstack browse server cannot start, and Naver fetches fail before any page is read.
• Typical symptoms: Server failed to start, No available port after 5 attempts, or direct listen EPERM: operation not permitted 127.0.0.1:<port>.
• Impact: kr-naver-browse, kr-naver-blogger, and kr-naver-insight may appear to return empty results even though the target posts exist.
• Current workaround: rerun the browse-backed command outside the sandbox or with elevated execution so the local server can bind to a port. This is an environment/runtime issue, not a Naver-only parsing issue.