mirror of https://github.com/NomaDamas/k-skill.git synced 2026-09-14 16:37:18 +08:00

Files

T

Jeffrey (Dongkyu) Kim 701b6f1c81 Release dev changes to main (#412 )

* feat(daangn-realty-search): expose listing release_date from detail JSON-LD

Product.releaseDate is available on the article detail page but wasn't
extracted. Surface it as release_date on --titles-enriched search items
so agents can tell how recently a listing was posted/refreshed.

* docs: update CLAUDE.md proxy deploy section for gpu01 cron+Docker

Production deploy already moved from Cloud Run/GitHub Actions to gpu01
cron-triggered Docker deploys (docs/deploy-k-skill-proxy.md), but
CLAUDE.md still described the old Cloud Run flow.

* feat(iros-registry-automation): local registry PDF summary

발급받은 등기부등본 PDF에서 근저당·가압류·소유권 변동 등 핵심 항목을 사실만 뽑아
사람이 읽기 쉬운 요약을 만든다. 발급 후 단계 전용(로그인·결제 재실행 없음).

- scripts/iros_pdf_summary.py: 로컬 텍스트 추출(pdfplumber→pypdf) 기본, 원격 전송 없음.
  스캔 PDF는 사용자가 로컬 self-host 변환기(marker 등)로 바꾼 결과를 --from-text로 요약.
  주민등록번호 마스킹, 요약은 $workdir/output에만(판단·법률자문 없음).
- tests/test_iros_pdf_summary.py: 파서 순수함수 유닛테스트 12개(모의 텍스트, 실 PDF·PII 없음).
- SKILL.md: "7. 발급 후 PDF 요약(선택)" 워크플로 + Response policy(판단형 표현 금지) + marker 크레딧.
- docs/features 업데이트.

marker 참고: NomaDamas/marker-api-server (로컬 self-host, 이 스킬이 원격 전송하지 않음).

* feat(naver-ad-performance): add read-only Naver SearchAd performance skill

New Python-script skill wrapping the Naver SearchAd (검색광고) API:
campaign/adgroup/keyword listing, stats (impressions/clicks/spend/
CTR/CPC/avg rank/conversions), keyword tool (related keywords, monthly
search volume, competition), and a doctor command for credential/
reachability diagnostics.

Read-only by design — no bid, campaign, or keyword mutation endpoints
are implemented, matching the pattern used by other credentialed
k-skill skills (korean-patent-search, k-dart). Auth follows the
official HMAC-SHA256 request signing scheme (X-Timestamp/X-API-KEY/
X-Customer/X-Signature), stdlib-only (hmac/hashlib/base64/urllib), no
new dependencies.

Wired into README skill table, docs/features, docs/security-and-secrets.md
default env list, and docs/install.md's setup-required skill list.

Reference: https://github.com/NariP/naver-searchad (MIT) for the
signing scheme and command surface.

* Harden k-skill-proxy public docs (#394)

* Fix iros PDF summary Python compatibility

* Fix Naver ad signature coverage

* Sanitize proxy deploy guidance

* Fix daangn release date extraction

* Wire iros PDF summary tests into CI

* Wire Naver ad tests into CI

* Sanitize public proxy operation docs

* BrowserOS-first browser runtime (auto default) + migrate browser skills (#395)

* Fix S2B live list row parsing

* Fix D2B public page classification

* Fix yebigun BrowserOS session handling

* Fix D2B mixed bid category parsing

* Fix browser runtime disconnect cleanup

* Add changeset for live skill fixes

* Fix yebigun open menu cleanup tests

* Document browser runtime guidance

* Fix D2B public small bid field parsing

* Sanitize yebigun live verification docs

* Default browser runtime to BrowserOS-first auto and migrate court/hipass skills

- Migrate hipass-receipt, court-auction-notice-search, and court-payment-order-assistant
  onto k-skill-browser-runtime (semver dependency, CDP-only, disconnect-not-close).
- Default the browser runtime provider to `auto`: prefer a user-launched BrowserOS
  session (9100) and fall back to a Chrome/Chromium CDP session (9222);
  KSKILL_BROWSER_PROVIDER can force browseros/chrome-cdp, --cdp-url targets a URL.
- court-auction keeps direct HTTP first with a runtime-CDP -> local-launch fallback;
  court-payment stays BrowserOS CDP -> manual handoff only.
- Document that court-auction and court-payment produce their core output without a
  real browser; keep login/CAPTCHA/payment/e-signature/irreversible boundaries manual.
- Update docs (setup, adding-a-skill, feature guides), READMEs, changeset, lockfile,
  and the skill-docs regression test.

* Document browser runtime skill authoring

* Prefer Aside before Chrome CDP

* Add KR WHOIS lookup skill

* Add NHIS care search skill

* Add KOPIS performance search skill

* Add Assembly bill vote search skill

* Add Korean holiday calendar skill

* Add NHIS checkup proxy routes

* fix(browser-runtime): clean up failed sessions

* fix(proxy): avoid caching semantic errors

* fix(proxy): restore Cloud Run release path

* fix(proxy): make Cloud Run image reproducible

* fix(proxy): harden public deployment boundaries

* fix(proxy): redact passthrough credentials

* fix(proxy): sanitize AirKorea fetch failures

* fix(proxy): bind rollout to tested revision

* fix(proxy): close operator-only surfaces

* docs(proxy): fail closed during manual rollout

* fix(proxy): harden deployment identity and client limits

* docs(proxy): align security release notes

* fix(proxy): use direct workflow OIDC claim

* docs: justify cohesive oversized adapters

* feat: complete WHOIS lookups and prefer Aside on macOS

* fix(proxy): harden live release routes

* fix(proxy): contain WHOIS fetch failures

---------

Co-authored-by: bangju <redbear.photo@gmail.com>

2026-07-12 11:49:23 +09:00

4.3 KiB

Raw Permalink Blame History

k-skill-browser-runtime

Browser runtime adapter for k-skill packages, with platform-aware automatic provider selection.

Scope

The recommended default is auto: macOS prefers Aside Browser, then BrowserOS, then Chrome/Chromium CDP. Other platforms preserve BrowserOS, Aside Browser, then Chrome/Chromium CDP.
BrowserOS is used as a GUI/session browser over CDP, not as a headless backend, CAPTCHA bypass, login solver, payment solver, or stealth scraping browser.
The browseros provider connects to a user-launched BrowserOS session. It never launches BrowserOS and never passes headless flags.
The aside provider uses the public aside repl CLI surface. It does not rely on undocumented local ports or a private CDP endpoint.
Site-specific navigation, parsing, and irreversible-boundary decisions stay in each skill.

Providers

Provider	Default surface	Launches browser	Intended use
`auto` (default)	macOS: `aside repl`, BrowserOS `9100`, Chrome `9222`; other platforms: BrowserOS, Aside, Chrome	No	Platform-aware fallback order
`browseros`	`http://127.0.0.1:9100`	No	Force a user-launched BrowserOS GUI/session browser
`aside`	`aside repl`	No	Force an available Aside Browser session through the documented CLI REPL
`chrome-cdp`	`http://127.0.0.1:9222`	No	Force a Chrome/Chromium CDP session
Unknown provider names fail closed with a typed `UNKNOWN_PROVIDER` error rather than silently falling back to BrowserOS.

Environment

KSKILL_BROWSER_PROVIDER selects auto (default), browseros, aside, or chrome-cdp.
KSKILL_BROWSEROS_CDP_URL overrides the BrowserOS CDP URL.
KSKILL_CHROME_CDP_URL overrides the Chrome CDP URL.
KSKILL_ASIDE_COMMAND overrides the Aside CLI command (default aside).

Lifecycle

For BrowserOS CDP and Chrome CDP sessions, disconnect automation clients instead of closing the browser application or persistent user profile. For Aside Browser sessions, close only tabs created by the adapter. Pre-existing user pages must not be closed.

Connect options

connect(options) selects and connects to a CDP endpoint. It supports dependency injection for tests:

options.probe — false skips the CDP health probe; a function replaces the default probeCdp probe.
options.connectLoader — function (cdpUrl, options) => browser replacing the default connectOverCDP path (which lazily loads playwright-core/playwright/rebrowser-playwright).
options.chromiumLoader — function passed through to the default CDP loader for lazy chromium resolution and caching.

connect() probes CDP providers with <cdpUrl>/json/version before connecting. For Aside, it runs a safe aside repl probe that lists tabs without opening a page. Provider failures throw UNAVAILABLE; unknown providers throw UNKNOWN_PROVIDER.

Stop rules

The runtime exports typed stop reasons for manual handoff boundaries: authentication, CAPTCHA, payment, electronic signature, irreversible submit, blocked upstream responses, and provider unavailability.

Job runner

runJob({ url, steps, stopOn }) is a narrow declared-step runner. It executes ONLY caller-supplied step functions in declared order and never invents navigation, generates steps, or plans site behavior.

steps — array of caller-supplied async functions ({ page, results }) => result. Empty/non-array steps returns { status: "no-steps", results: [] } without navigating.
url — optional. When a URL is supplied AND at least one caller step exists, the runner performs a single page.goto(url, { waitUntil: "domcontentloaded" }) before the steps. No URL means no navigation.
page — caller-supplied automation page (for example from getAutomationPage). Required only when url is supplied.
stopOn — optional async callback ({ page, phase, url, step, index, results }) => stopReason | null invoked at phase boundaries: once before navigation (phase: "navigate") and once before each step (phase: "step", with index). A truthy stopReason short-circuits to { status: "stopped", stopReason, results } without bypassing manual handoff boundaries.

Returns { status, results } where status is "no-steps", "stopped", or "complete". Non-function steps throw a TypeError before any navigation or step execution.

4.3 KiB Raw Permalink Blame History