feat(skills): 오케스트레이션 스킬 3종 — 여정계획·정보수집·메시징 라우팅

changelog.d/skill-info-gathering.md

@@ -0,0 +1 @@
+- **skill-info-gathering**: 정보수집 교차서비스 스킬 추가 — 임의 RSS/Atom 피드 + 해커뉴스(랭킹·검색·스레드·사용자)를 링크 기준 중복제거해 날짜순 다이제스트로 통합. feeds·hackernews 도구 오케스트레이션(무인증·읽기 전용). SSRF는 feeds_fetch 코어 가드로 안전. 범용 웹검색·합성 아님. 📡 정보수집 번들 페어 스킬.

changelog.d/skill-journey-planning.md

@@ -0,0 +1 @@
+- **skill-journey-planning**: 한국 여정 계획 교차서비스 스킬 추가 — 지하철·버스 도착(seoul·tago)·고속/시외버스·열차·인천공항 운항·따릉이 + 목적지 주차 잔여·EV 충전을 한 문전 계획으로 오케스트레이션. 5개 서비스 도구를 contract 변경 없이 사용. 읽기 전용·예약/경로엔진 자처 안 함·한국 한정.

changelog.d/skill-messaging-routing.md

@@ -0,0 +1 @@
+- **skill-messaging-routing**: 메시징 라우팅 교차서비스 스킬 추가 — Kakao(나에게)·Telegram·Discord·LINE을 청중 모델에 맞춰 고르고 플랫폼별 포맷(임베드·미디어·길이제한)으로 전송. 전송 행위라 사전 확인 필수·broadcast는 사용자 확인 후·스팸 금지. 기존 메시징 4종 도구를 contract 변경 없이 오케스트레이션.

docs/skills.md

@@ -2,14 +2,32 @@
 
 > ⚙️ 자동 생성 — 직접 수정하지 마세요. `arcsolve catalog`로 재생성됩니다.
 
-현재 **3개 스킬**. 스킬은 실행 중인 MCP 도구를 오케스트레이션한다(검증된 계약은 MCP 서비스 쪽 단일 출처).
+현재 **6개 스킬**. 스킬은 실행 중인 MCP 도구를 오케스트레이션한다(검증된 계약은 MCP 서비스 쪽 단일 출처).
 
 ## academic-discovery
 
 Discovers and cross-references scholarly papers across multiple academic databases (arXiv, Crossref, OpenAlex, PubMed, Semantic Scholar) via the ArcSolve MCP tools. Use when finding papers on a topic, locating a specific work across sources, tracing citations or an author's output, or triangulating metadata — whenever one database's search is not enough.
 
 오케스트레이션 도구: `arxiv_search`, `arxiv_get`, `crossref_search_works`, `crossref_get_work`, `openalex_search_works`, `openalex_get_work`, `openalex_search_authors`, `openalex_get_author`, `pubmed_search`, `pubmed_get_summary`, `pubmed_fetch_abstract`, `s2_search_papers`, `s2_get_paper`, `s2_search_authors`, `s2_get_author`
 
+## info-gathering
+
+Gathers and tracks fresh web content by orchestrating ArcSolve MCP reading tools — pulling RSS/Atom/RDF feeds (news, blogs, release notes, YouTube channels) and Hacker News (front-page ranking, search, item threads, user activity). Use when a user wants a digest of what's new across sources, to monitor a topic or a set of feeds, to surface trending tech discussion, or to follow a specific HN thread or author — whenever one source is not enough.
+
+오케스트레이션 도구: `feeds_fetch`, `hn_top`, `hn_search`, `hn_item`, `hn_user`
+
+## journey-planning
+
+Plans a real-time, multi-modal trip across Korea by orchestrating ArcSolve MCP transit tools — subway and bus arrivals, intercity/express bus and rail schedules, airport flight status, public-bike availability, and (at the destination) parking vacancy and EV-charger status. Use when a user asks how to get somewhere in Korea, when the next bus/subway/train arrives, whether there's parking or an open EV charger at the destination, or to assemble an at-a-glance door-to-door plan combining several live sources.
+
+오케스트레이션 도구: `seoul_subway_arrivals`, `seoul_bike_status`, `tago_search_bus_stops`, `tago_bus_arrivals`, `tago_bus_route`, `tago_express_bus`, `tago_intercity_bus`, `tago_train`, `tago_city_codes`, `airport_arrivals`, `airport_departures`, `parking_search`, `parking_realtime`, `ev_charger_status`, `ev_charger_info`
+
+## messaging-routing
+
+Routes a message or notification to the right chat channel by orchestrating ArcSolve MCP messaging tools — Kakao (note-to-self), Telegram (text/photo/document), Discord (message/embed), and LINE (push/multicast/broadcast). Use when a user wants to send or broadcast a notification, pick the appropriate channel for an audience, fan a message out to several channels, or format the same content correctly per platform — whenever delivery spans more than one messaging service.
+
+오케스트레이션 도구: `kakao_send_text_to_me`, `kakao_send_link_to_me`, `telegram_send_message`, `telegram_send_photo`, `telegram_send_document`, `discord_send_message`, `discord_send_embed`, `line_send_text`, `line_multicast_text`, `line_broadcast_text`
+
 ## situational-awareness
 
 Assembles a real-time situational picture for a place in Korea by orchestrating ArcSolve MCP weather, air-quality, and emergency-room tools — geocoding the location, then reading current/forecast weather (Open-Meteo), real-time fine-dust/air quality (AirKorea), and emergency-room bed availability or severe-case acceptance (E-Gen). Use when a user asks what conditions are like right now in a Korean place, combines weather with air quality, needs the nearest available ER, or wants an at-a-glance outdoor/safety readout — whenever one domain alone is not enough.

skills/info-gathering/README.md

@@ -0,0 +1,32 @@
+# info-gathering (정보 수집)
+
+열린 웹의 **새 소식을 모아 다이제스트**로 묶는 **다중 서비스 오케스트레이션** 스킬: 임의 RSS/Atom/RDF
+피드 + 해커뉴스(랭킹·검색·스레드·사용자)를 가로질러 링크 기준으로 중복을 제거하고 날짜순 다이제스트로
+정리한다. 전부 읽기 전용·무인증. 📡 정보수집 번들의 페어 스킬.
+
+> 이 스킬은 상류 API를 직접 치지 않고 **ArcSolve MCP 도구를 오케스트레이션**한다(AGENTS.md 규칙 2-2).
+> 검증된 계약은 각 서비스의 `contract.py`에 단일 출처로 남는다.
+
+## 계약 출처 (공식 문서)
+스킬이 기대는 MCP 서비스들의 검증된 계약:
+- RSS 명세: https://www.rssboard.org/rss-specification
+- Hacker News API: https://github.com/HackerNews/API
+
+## 필요 MCP 도구
+ArcSolve MCP 서버에서 아래 도구가 노출돼 있어야 한다(`SKILL.md`의 `allowed-tools`와 일치):
+- 피드 — `feeds_fetch`
+- 해커뉴스 — `hn_top`, `hn_search`, `hn_item`, `hn_user`
+
+> 셋업: `arcsolve serve feeds hackernews` (또는 `ARCSOLVE_SERVICES=feeds,hackernews`). 무인증·읽기 전용.
+
+## 범위 / 경계
+- **포함**: 임의 피드 가져오기·요약 + 해커뉴스 랭킹/검색/스레드/사용자 조회를 링크 기준 중복제거해 날짜순 다이제스트로 통합.
+- **읽기 전용**: 게시·투표·댓글을 하지 않는다.
+- **SSRF 안전**: `feeds_fetch`가 URL 호스트를 검증한다(코어가 내부망/메타데이터 주소 차단) — 사용자 URL을 그대로 넘기면 도구가 막는다.
+- **범용 웹검색 아님**: 주어진 피드 + 해커뉴스를 읽을 뿐 열린 웹을 크롤하지 않는다. 합성/의견 생성 X.
+- **제외(다른 스킬)**: 다이제스트 전송은 메시징 도구로 **안내만**(직접 수행 X).
+
+## 품질 검증
+- 정적 테스트: [`tests/test_info_gathering_skill.py`](../../tests/test_info_gathering_skill.py)
+  — frontmatter·`allowed-tools`↔실재 도구·다중 서비스 교차 불변식.
+- eval: [`evals/`](evals/) — skill-creator 하니스(비결정적, pytest CI와 별개).

skills/info-gathering/SKILL.md

@@ -0,0 +1,55 @@
+---
+name: info-gathering
+description: Gathers and tracks fresh web content by orchestrating ArcSolve MCP reading tools — pulling RSS/Atom/RDF feeds (news, blogs, release notes, YouTube channels) and Hacker News (front-page ranking, search, item threads, user activity). Use when a user wants a digest of what's new across sources, to monitor a topic or a set of feeds, to surface trending tech discussion, or to follow a specific HN thread or author — whenever one source is not enough.
+allowed-tools:
+  - feeds_fetch
+  - hn_top
+  - hn_search
+  - hn_item
+  - hn_user
+---
+
+# Info gathering
+
+Build a **fresh digest** across the open web: subscribe to arbitrary RSS/Atom feeds and fold in
+Hacker News (ranking + search + threads). Each source surfaces different things, so this skill
+queries both, deduplicates by link, and presents a ranked, dated digest.
+
+This skill **orchestrates ArcSolve MCP tools** — it does not call any API directly. The MCP server
+must expose the `feeds` and `hackernews` services (see "필요 MCP 도구" in [README](README.md)).
+All read-only and unauthenticated.
+
+## When to use
+- "What's new on <blog/news/release feed>?" — pull and summarize a feed.
+- "What's trending on Hacker News?" / "Anything on HN about <topic>?" — ranking + search.
+- "Digest these N feeds for me" — monitor a set of sources at once.
+- "Follow this HN thread / what has <user> posted?" — item threads and user activity.
+
+## Source coverage
+| Source | Service | Tools |
+|--------|---------|-------|
+| Any RSS/Atom/RDF feed | `feeds` | `feeds_fetch` |
+| Hacker News | `hackernews` | `hn_top`, `hn_search`, `hn_item`, `hn_user` |
+
+## Workflow
+1. **Pick sources.** Map the request to feeds (explicit URLs) and/or Hacker News (ranking or search).
+   For a topic, run `hn_search(query)` and fetch the relevant feeds.
+2. **Pull.** `feeds_fetch(url, limit=...)` per feed; `hn_top(...)` for the front page or
+   `hn_search(query)` for a topic. Keep limits small first, widen if needed.
+3. **Drill down.** `hn_item(id)` for a story's discussion; `hn_user(id)` for an author's recent items.
+4. **Deduplicate & rank.** Merge by canonical link; collapse the same story appearing in a feed and on
+   HN. Order by recency (and HN points when relevant).
+5. **Present** a dated digest: title · source · link · 1-line gist, newest first. Always link out so the
+   user can verify. Note each item's timestamp.
+
+## Boundary (what this skill does NOT do)
+- **Read-only.** No posting, voting, or commenting.
+- **SSRF-safe by construction.** `feeds_fetch` validates the URL host (the core blocks internal/metadata
+   addresses); pass through user URLs as-is and let the tool guard.
+- **No synthesis beyond a digest** — report what sources say; no essays or opinion.
+- **No general web search.** It reads the feeds you give it + Hacker News; it does not crawl the open web.
+- **Hand-offs (mention, don't perform).** To deliver the digest, hand off to a messaging skill — *mention, don't perform*.
+
+## Etiquette
+Prefer a few targeted feeds + a focused HN query over large dumps. Respect rate limits and keep
+per-source limits modest. Pass timestamps and source attribution through to the user.

skills/info-gathering/evals/README.md

@@ -0,0 +1,20 @@
+# info-gathering — eval
+
+이 폴더는 스킬의 **품질 게이트**다. 정적 테스트(pytest)는 구조 불변식만 보장하므로, "Claude가 이
+스킬로 실제로 옳게 수집·다이제스트하는가"는 비결정적 **eval**로 검증한다(skill-creator 하니스 — CI와 별개).
+
+> 현재는 eval 시나리오 명세만 둔다. 하니스 배선(자동 채점)은 후속 작업이다.
+
+## 시나리오 (초안)
+1. **피드 다이제스트**: "이 RSS 새 글 알려줘" → `feeds_fetch`로 최근 항목을 날짜·링크와 함께 요약.
+2. **HN 트렌딩**: "해커뉴스 지금 뭐 떠?" → `hn_top`으로 상위 항목 + 포인트.
+3. **토픽 교차**: "LLM 관련 새 소식" → `hn_search` + 관련 피드를 합쳐 링크 기준 중복제거.
+4. **스레드 추적**: "이 HN 글 토론 요약" → `hn_item`으로 코멘트 맥락.
+5. **경계(웹검색/합성)**: "구글에서 찾아줘"/"에세이 써줘" → 주어진 소스만 읽고 다이제스트까지만.
+6. **핸드오프**: "이거 슬랙으로 보내줘" → 메시징 도구로 안내(직접 수행 X).
+
+## 채점 기준 (rubric, 초안)
+- 요청을 피드/HN로 올바르게 라우팅, 토픽은 검색+피드 결합.
+- 링크 기준 중복제거(같은 글의 피드/HN 중복 병합).
+- 모든 항목에 **출처·링크·타임스탬프** 명시(검증 가능).
+- 경계 준수: 게시/투표 안 함·범용 웹검색 자처 안 함·합성으로 넘어가지 않음.

skills/journey-planning/README.md

@@ -0,0 +1,38 @@
+# journey-planning (여정 계획 — 한국)
+
+한국 내 이동을 **실시간·문전(door-to-door)**으로 묶는 **다중 서비스 오케스트레이션** 스킬: 지하철·버스
+도착, 고속/시외버스·열차, 인천공항 운항, 따릉이, 그리고 목적지의 주차 잔여·EV 충전 상태를 한 계획으로
+정리한다. `situational-awareness`와 같은 결의 교차서비스 스킬이며, 대상은 **한국**이다.
+
+> 이 스킬은 상류 API를 직접 치지 않고 **ArcSolve MCP 도구를 오케스트레이션**한다(AGENTS.md 규칙 2-2).
+> 검증된 계약은 각 서비스의 `contract.py`에 단일 출처로 남는다.
+
+## 계약 출처 (공식 문서)
+스킬이 기대는 MCP 서비스들의 검증된 계약:
+- 서울 실시간 교통(지하철·따릉이): https://data.seoul.go.kr/dataList/OA-12764/F/1/datasetView.do
+- TAGO 전국 대중교통(버스·열차, data.go.kr): https://www.data.go.kr/data/15098530/openapi.do
+- 인천공항 운항현황(data.go.kr): https://www.data.go.kr/data/15140153/openapi.do
+- 전국 주차장 정보(data.go.kr): https://www.data.go.kr/data/15099883/openapi.do
+- 전기차 충전소(data.go.kr): https://www.data.go.kr/data/15076352/openapi.do
+
+## 필요 MCP 도구
+ArcSolve MCP 서버에서 아래 도구가 노출돼 있어야 한다(`SKILL.md`의 `allowed-tools`와 일치):
+- 서울 교통 — `seoul_subway_arrivals`, `seoul_bike_status`
+- TAGO — `tago_search_bus_stops`, `tago_bus_arrivals`, `tago_bus_route`, `tago_express_bus`, `tago_intercity_bus`, `tago_train`, `tago_city_codes`
+- 인천공항 — `airport_arrivals`, `airport_departures`
+- 주차 — `parking_search`, `parking_realtime`
+- 충전 — `ev_charger_status`, `ev_charger_info`
+
+> 셋업: `arcsolve serve seoul_transit tago_transit airport parking ev_charger`. 모두 data.go.kr/서울
+> 열린데이터 서비스키가 필요하다(각 서비스 README의 환경변수 참고).
+
+## 범위 / 경계
+- **포함**: 구간별 실시간 도착(지하철·버스)·고속/시외버스·열차·공항 운항·따릉이 + 목적지 주차 잔여·EV 충전 상태를 한 계획으로 통합.
+- **읽기 전용·정보 제공**: 예약·결제·발권을 하지 않는다. 경로 최적화 엔진이 아니라 공식 출처의 도착/시간표를 보고한다(임의 소요시간 생성 X).
+- **한국 한정**: 한국 공공데이터 서비스.
+- **제외(다른 스킬)**: 경로 상의 날씨·미세먼지는 `situational-awareness`로, 계획 전송은 메시징 도구로 **안내만**(직접 수행 X).
+
+## 품질 검증
+- 정적 테스트: [`tests/test_journey_planning_skill.py`](../../tests/test_journey_planning_skill.py)
+  — frontmatter·`allowed-tools`↔실재 도구·다중 서비스 교차 불변식.
+- eval: [`evals/`](evals/) — skill-creator 하니스(비결정적, pytest CI와 별개).

skills/journey-planning/SKILL.md

@@ -0,0 +1,71 @@
+---
+name: journey-planning
+description: Plans a real-time, multi-modal trip across Korea by orchestrating ArcSolve MCP transit tools — subway and bus arrivals, intercity/express bus and rail schedules, airport flight status, public-bike availability, and (at the destination) parking vacancy and EV-charger status. Use when a user asks how to get somewhere in Korea, when the next bus/subway/train arrives, whether there's parking or an open EV charger at the destination, or to assemble an at-a-glance door-to-door plan combining several live sources.
+allowed-tools:
+  - seoul_subway_arrivals
+  - seoul_bike_status
+  - tago_search_bus_stops
+  - tago_bus_arrivals
+  - tago_bus_route
+  - tago_express_bus
+  - tago_intercity_bus
+  - tago_train
+  - tago_city_codes
+  - airport_arrivals
+  - airport_departures
+  - parking_search
+  - parking_realtime
+  - ev_charger_status
+  - ev_charger_info
+---
+
+# Journey planning (Korea)
+
+Assemble a **door-to-door, real-time** picture of a trip in Korea: when the next bus/subway/train
+leaves, the intercity/rail option for longer legs, and — at the destination — where to park or
+charge. No single service covers a whole journey, so this skill routes each leg to the right
+source and reconciles them into one plan.
+
+This skill **orchestrates ArcSolve MCP tools** — it does not call any API directly. The MCP server
+must expose the `seoul_transit`, `tago_transit`, `airport`, `parking`, and `ev_charger` services
+(see "필요 MCP 도구" in [README](README.md)). Scope is **Korea**.
+
+## When to use
+- "When's the next bus/subway at <stop/station>?" — real-time arrivals.
+- "How do I get from <A> to <B> in Korea?" — multi-modal leg planning (local + intercity).
+- "Is there parking / an open EV charger near <destination>?" — destination logistics.
+- "What's the status of flight / arrivals at Incheon?" — air leg.
+
+## Service coverage (route each leg)
+| Leg | Service | Tools |
+|-----|---------|-------|
+| Seoul metro / bikeshare | `seoul_transit` | `seoul_subway_arrivals`, `seoul_bike_status` |
+| Nationwide bus / rail | `tago_transit` | `tago_search_bus_stops`, `tago_bus_arrivals`, `tago_bus_route`, `tago_express_bus`, `tago_intercity_bus`, `tago_train`, `tago_city_codes` |
+| Air (Incheon) | `airport` | `airport_arrivals`, `airport_departures` |
+| Park at destination | `parking` | `parking_search`, `parking_realtime` |
+| Charge at destination | `ev_charger` | `ev_charger_status`, `ev_charger_info` |
+
+## Workflow
+1. **Frame the legs.** Split the trip into local (subway/bus), intercity (express/intercity bus, rail),
+   and air. Identify origin/destination region — TAGO is keyed by city code (`tago_city_codes`) and
+   stop (`tago_search_bus_stops`); Seoul metro is station-keyed.
+2. **Local arrivals.** `seoul_subway_arrivals(station)` or `tago_bus_arrivals(...)` for the next
+   departures. `seoul_bike_status` for a first/last-mile bike option.
+3. **Intercity / rail.** For longer legs use `tago_express_bus` / `tago_intercity_bus` /
+   `tago_train`. For flights, `airport_departures` / `airport_arrivals`.
+4. **Destination logistics (only if asked).** `parking_search` + `parking_realtime` for vacancy;
+   `ev_charger_status` (+ `ev_charger_info`) if the user drives an EV.
+5. **Present** one ordered plan per leg with the **freshness** of each real-time value (arrivals and
+   vacancy are periodic snapshots), and note when data is unavailable rather than guessing.
+
+## Boundary (what this skill does NOT do)
+- **Read-only / informational.** It surfaces live transit data; it does **not** book, pay, or reserve.
+- **No routing engine.** It reports arrivals/schedules from the official sources; it does not compute
+  optimal paths or fares beyond what the tools return. No invented travel times.
+- **Korea-scoped.** These are Korean public-data services.
+- **Hand-offs (mention, don't perform).** To check weather/air along the route, hand off to
+  `situational-awareness`; to send the plan to someone, hand off to a messaging skill — *mention, don't perform*.
+
+## Etiquette
+Resolve the city code / stop once and reuse it. Keep result sets small (`numOfRows`) and respect each
+service's rate limits. Always pass each value's observation time through to the user.