Hotel Lookup (FlyAI · Fliggy MCP)

Provide hotel discovery, comparison, and booking handoff using the FlyAI CLI command search-hotel, backed by Fliggy MCP. Outputs are decision-ready when you respect live JSON fields and the flag schema in references/search-hotel.md (single source of truth).

Prerequisites

1. Install FlyAI CLI: npm i -g @fly-ai/flyai-cli
2. Optional richer results: flyai config set FLYAI_API_KEY "your-key" (keep secrets out of logs).
3. I/O: one JSON object per line on stdout; guidance/errors on stderr (flyai search-hotel --help for flag discovery).

Workflow

1. Capture intent before running search-hotel

Extract and confirm (ask only the minimum if missing):

• Destination → maps to required --dest-name (country / province / city / district).
• Stay window → --check-in-date / --check-out-date (YYYY-MM-DD).
• Budget → --max-price (CNY per night cap).
• Stars & beds → --hotel-stars, --hotel-bed-types.
• Lodging mix → --hotel-types (酒店 / 民宿 / 客栈).
• Landmark / “near X” → --poi-name plus --dest-name.
• Free-text narrowing → --key-words.
• Ranking preference → --sort (price_asc / price_desc / rate_desc / distance_asc / no_rank).
• Trip purpose (business / family / leisure) → use to choose filters and how you explain trade-offs; do not send fields that references/search-hotel.md does not define.

2. Align parameters to the reference (no “tag priming” RPC)

• Open references/search-hotel.md and mirror exact flag names—there is no separate “prime tags” or tag-cache step in search-hotel.
• Validate dates (not in the past; correct format). If the city or district is ambiguous, ask the user—never guess --dest-name.

3. Search hotels with normalized flyai search-hotel flags

• Build one invocation: flyai search-hotel … per the reference Examples and parameter list.
• Prefer a bounded result set for first pass (e.g. sensible defaults from user intent); shortlist to top 3–5 in the final narrative.
• Respect live JSON behavior from upstream:
• Some fields may be null or missing.
• Price in samples appears as display strings (e.g. ¥618); do not invent currency rules not present in data.
• If the service returns errors or empty itemList, follow Error handling below.

4. Enrich finalists from returned JSON only (no extra hotel-detail RPC)

• Each finalist row comes from data.itemList (see output example in references/search-hotel.md).
• Use mainPic, detailUrl, name, address, price, score, star, review, interestsPoi, etc. when present.
• There is no getHotelDetail call in this skill. Deeper room-level matrices like full rate-plan grids are only available if they appear inside the JSON you already received; otherwise direct the user to detailUrl for authoritative booking pages.
• If you need a second pass, run a narrower search-hotel (tighter dates, POI, or keywords)—do not fabricate APIs.

5. Return decision-ready output

Always give:

• Recommended pick (best fit to stated constraints).
• Two alternatives with honest trade-offs (price vs distance vs stars vs POI proximity—only from observed fields).
• Booking handoff: what to open next (detailUrl), what to double-check on the supplier page, and 2–4 final confirmation questions if anything is still ambiguous.

Output template (concise bullets)

• 行程信息: 目的地 / 入住离店 / 人数或房型需求 / 预算（每晚 CNY）/ 关键偏好（商务/亲子等）
• 推荐酒店（首选）
• 酒店名 · 价格展示（来自 JSON）
• 位置 / 交通或 POI 相关字段（如 interestsPoi）
• 图片行：![]({mainPic})（若存在）
• 详情/预订：Click to book（若存在）
• 评分/星级/短评（若存在）
• 推荐理由（只引用结果中可见事实）
• 备选 1 / 备选 2（同结构，字段缺失则说明）
• 决策建议: 适合人群与取舍（不编造政策）
• 下一步确认: 仅列 2–4 个必要确认项（支付前核对官方页等）

Output & presentation (Markdown)

• Show image before booking link when both exist.
• Use headings and tables when comparing multiple hotels.
• Optional brand line: e.g. “Based on fly.ai real-time hotel results.”

Quality bar

• Prefer concrete numbers and fields from JSON over vague adjectives.
• Do not invent cancellation rules, breakfast, or prices not shown in data.
• If data is missing, stale, or mismatched to user intent, say so and suggest adjusting flags or confirming on detailUrl.
• Keep shortlists tight—avoid dumping huge itemList tables unchanged.
• Never expose API keys or local config in chat.

References

Error handling

1. Validate inputs against references/search-hotel.md; ask the user when required fields are unclear.
2. Diagnose stderr + JSON status / message; fix flags vs retry on transient errors only.
3. Empty results — relax filters once, then explain; no infinite retries.
4. Transparency — if location or dates may be wrong, state it before recommending payment.

Differentiation (honest scope)