agentsclimarketplace

Didi ride

Skill AceDataCloud/Skills/skills/didi-ride

通过滴滴出行 MCP 打车、查价、查询订单、取消订单、规划路线(驾车/公交/步行/骑行)和搜索地点。Use when the user mentions 滴滴, 打车, 叫车, 回家/上班要打车, 查一下从 A 到 B 多少钱/怎么走, 查询订单, 司机在哪/多久到, 预约叫车, 路线规划, DiDi, ride-hailing, or booking a taxi.From its SKILL.md

Install
npx -y skills add AceDataCloud/Skills --skill didi-ride

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
  • 13 stars13 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as Apache-2.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

6.5 KB, ~1.6k tokens by cl100k_base, as published. Nobody here has run it

DiDi Ride (滴滴出行)

Drive the user's DiDi (滴滴出行) account through the DiDi MCP server: book rides, estimate fares, track orders, cancel, and plan routes.

The didi BYOC connector injects one env var into the sandbox:

  • DIDI_MCP_KEY — the user's DiDi MCP key. Secret — never echo, print, or log it.

If DIDI_MCP_KEY is missing, tell the user to connect the DiDi connector at auth.acedata.cloud/user/connections (they get the key by scanning the QR in the 滴滴出行 App or via https://mcp.didichuxing.com/claw).

CLI

The skill ships a stdlib-only helper that speaks the MCP Streamable-HTTP protocol to DiDi. Run this resolver at the top of every Bash block below — each Bash call is a fresh shell, and $SKILL_DIR points at the LAST skill loaded this turn, so anchor on our own script:

DIDI="$SKILL_DIR/scripts/didi.py"; [ -f "$DIDI" ] || DIDI=$(find /tmp -maxdepth 8 -path '*/skills/*/scripts/didi.py' 2>/dev/null | head -1)
[ -f "$DIDI" ] || { echo "didi-ride script not found (SKILL_DIR=$SKILL_DIR)" >&2; exit 1; }

Two commands:

python3 $DIDI list                        # list tools + their JSON input schemas
python3 $DIDI call <tool> '<json-args>'   # call any tool

Always check references/api_references.md for the exact tool + parameter names before calling. When unsure, run python3 $DIDI list — it returns the authoritative input schema for every tool straight from DiDi. Do not guess parameter names (common mistakes: keyword → keywords, region → city, four coord fields → six from_name/from_lat/from_lng/to_name/to_lat/to_lng).

All argument values must be strings (including coordinates and product_category), e.g. {"product_category":"1"} not {"product_category":1}.

Write gating (real money / real driver)

taxi_create_order and taxi_cancel_order are gated: without a trailing --confirm the helper only DRY-RUNS and changes nothing. Run once without --confirm to preview, then re-run with --confirm as the last argument after the user approves:

python3 $DIDI call taxi_create_order '{"estimate_trace_id":"...","product_category":"1"}' --confirm
  • Even when the user says "打车" / "取消订单", confirm the concrete details (起终点、车型 for booking; the order for cancelling) before adding --confirm.
  • Cancel intent ≠ cancel confirmation — always ask "确认取消吗?" first.

Booking flow (最小可执行)

  1. Resolve addresses — maps_textsearch (never invent coordinates; don't reuse coordinates from earlier turns — the user may have moved).
    • If the user references an address alias (家 / 公司 / etc.) that you don't have, ask them; this skill has no stored preferences.
  2. Confirm start/end — if maps_textsearch returns ≥2 candidates, list at least the top 3 and let the user pick; a single exact match needs no confirmation.
  3. Estimate — taxi_estimate; record the returned traceId / estimate_trace_id. It expires (-32021) — re-estimate if stale.
  4. Pick car type — user's current message wins ("叫快车"→product_category 1, "专车"→8); otherwise ask. Only use categories present in the taxi_estimate response; never silently substitute a different service level (快车 1 ≠ 特惠快车 201).
  5. Create order — taxi_create_order with the latest estimate_trace_id (dry-run → confirm → --confirm).
  6. Report — order id, start/end, car type, estimated fare. Tell the user they can send 「查询订单」 to check status.

Query an order

Order id comes from (in priority): the user's message → the most recent order created this conversation → else ask.

python3 $DIDI call taxi_query_order '{"order_id":"ORDER_ID"}'

Status codes (code):

code含义输出
0匹配中⏳ 正在为您匹配司机
1司机已接单展示司机姓名、车型、车牌、电话、距离与预计到达时间
2司机已到达🔔 司机已到达上车点
4行程进行中🚗 行程已开始
5订单完成✅ 行程结束(展示费用,如有)
6系统取消❌ 订单已被系统取消
7已取消❌ 订单已取消
3 / 8-12其他终态显示对应状态描述

Routes & places (no order needed)

python3 $DIDI call maps_direction_driving  '{"from_name":"...","from_lat":"...","from_lng":"...","to_name":"...","to_lat":"...","to_lng":"..."}'
python3 $DIDI call maps_direction_transit  '{...}'   # 公交
python3 $DIDI call maps_direction_walking  '{...}'   # 步行
python3 $DIDI call maps_direction_bicycling '{...}'  # 骑行
python3 $DIDI call maps_place_around '{"keywords":"咖啡","lat":"...","lng":"..."}'
python3 $DIDI call maps_regeocode '{"lat":"...","lng":"..."}'

Gotchas

  • Never print the key or the raw endpoint URL — the helper handles the transport and keeps the key internal.
  • taxi_create_order takes only estimate_trace_id, product_category, and optional caller_car_phone. Don't pass the estimate's coordinate/name fields.
  • No stored preferences: this skill doesn't remember home/work/car type — ask the user, or wire richer memory in a higher layer.
  • Auth failure surfaces as -32002 (or HTTP 401/403): the key is missing/expired → have the user reconnect the connector.

Setup: See authentication. This skill uses the didi connector's injected DIDI_MCP_KEY, not ACEDATACLOUD_API_TOKEN.

What ships with it: 2 files

12.5 KB alongside SKILL.md, 1 of them executable

references/

scripts/

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.