120 lines
5.9 KiB
Markdown
120 lines
5.9 KiB
Markdown
---
|
|
name: web-task-router
|
|
description: Route website tasks by intent using a preset site registry, then decide whether to access public pages or login with saved credentials. Use when user asks to fetch/process data from websites, with or without login, especially market analysis tasks.
|
|
metadata:
|
|
{ "openclaw": { "emoji": "🌐", "os": ["linux", "darwin"] } }
|
|
---
|
|
|
|
# web-task-router Skill (OpenClaw)
|
|
|
|
Use this skill to choose the right website automatically and run login/public web tasks with human-in-the-loop verification.
|
|
|
|
## Registry file
|
|
|
|
- Read: `/root/.openclaw/workspace/websites.yaml`
|
|
- Use `routing_rules.hard_route_rules` first, then `sites[].purposes` + `sites[].keywords`.
|
|
|
|
## Decision flow
|
|
|
|
1. Parse user intent (what data/action is needed).
|
|
2. Check `routing_rules.hard_route_rules` first.
|
|
3. If user text matches any `when_keywords`, force `force_site_id`.
|
|
4. If no hard rule matched, choose one best site from `purposes/keywords`.
|
|
5. If no reliable match, ask one short clarification question.
|
|
6. If matched site has `login_required: false`, open target directly and proceed.
|
|
7. If `login_required: true`, resolve credentials, login, then execute task.
|
|
|
|
## Hard rule for market analysis
|
|
|
|
- If user request mentions market-analysis intent (for example: "市场分析", "行业分析", "类目分析", "竞品分析", "趋势分析"), always route to `ecommerce-market-analysis`.
|
|
- In this case, never use `web_search` or `web_fetch`.
|
|
- `browser` tool is disabled for market analysis.
|
|
- Preferred execution path: use `exec` with `yuce-browser open https://hq.yucekj.com` and `yuce-browser snapshot 60`.
|
|
- Do not output market-analysis metrics unless collected from yucekj pages in this run.
|
|
- If collection fails, ask the user for OTP/CAPTCHA/manual assist and pause.
|
|
|
|
## Hard rule for company sales / aftersale analysis
|
|
|
|
- If user request mentions company sales intent (for example: "销售数据", "业绩", "店铺业绩", "售后", "退款", "费比", "客服体验", "用户反馈", "话术分析", "产品信息", "SKU编码", "材质", "预览图"), always route to `company-sales-db-analysis`.
|
|
- This route must use SQLServer queries only. Do not open the old web dashboard for these tasks.
|
|
- Preferred execution path:
|
|
1) `exec`: `company-db health`
|
|
2) Product info query (`SKU/成本/材质/预览图`): `exec`: `product-info "<sku或关键词>" --days 30 --limit 20`
|
|
3) Product-ID SKU list query: `exec`: `product-skus-by-id "<商品id>" --limit 1000`
|
|
4) Sales/aftersale analysis: `exec`: `company-db run-file /root/.openclaw/workspace/sqlserver/last_week_afs_report.sql` (or targeted SQL query)
|
|
5) Customer-service experience/writing analysis: `exec`: `customer-service-analysis --start <YYYY-MM-DD> --end <YYYY-MM-DD> [--shop "店铺关键词"] --limit 20`
|
|
- For customer-service analysis outputs, include at least 2 dialogue snippets and recommended reply wording.
|
|
- For customer-service analysis outputs, include capability scoring by agent name (4 dimensions + overall grade).
|
|
- Before writing analysis, first read field catalog:
|
|
- `/root/.openclaw/workspace/outputs/company-sales-db-catalog.md`
|
|
- If a metric cannot be mapped to known tables/fields, return missing field list and ask one precise follow-up.
|
|
|
|
## Credential resolution order
|
|
|
|
1. If `credentials.source = env`, read `credentials.username_env` and `credentials.password_env` from environment variables.
|
|
2. If env is missing and `fallback_source = 1password`, use `username_ref/password_ref`.
|
|
3. If still unresolved, ask user for manual input.
|
|
|
|
## Login/verification policy
|
|
|
|
- Do not print raw password in chat.
|
|
- If OTP/CAPTCHA appears, pause and ask user for input.
|
|
- After user replies with code (or says challenge is done), resume from paused step.
|
|
|
|
## Safety rules
|
|
|
|
- Operate only inside `allowed_domains`.
|
|
- If redirected outside allowed domains, stop and ask user.
|
|
- Keep actions non-destructive unless user explicitly requests changes.
|
|
- Summarize final result briefly with data outcome and status.
|
|
|
|
## Autonomous module selection (yuce market-analysis)
|
|
|
|
Use this deterministic map before clicking pages:
|
|
|
|
- market size / growth / trend -> 市场洞察
|
|
- category share / fine-grained opportunities -> 叶子类目分析
|
|
- brand/shop/product competition -> 品牌店铺商品热榜, 竞店布局分析
|
|
- consumer demand / keyword changes -> 产品需求分析, 需求趋势分析, 关键词布局分析
|
|
- price strategy / price band -> 价格段布局, 品牌布局分析
|
|
|
|
Execution requirements:
|
|
|
|
1. Build a mini plan: requested metrics -> target modules (inside yuce only).
|
|
2. Open/login yuce, navigate by module names, capture snapshots as evidence.
|
|
3. If any metric lacks visible data, return "not available in current view" instead of guessing.
|
|
4. Never use `cron` for market-analysis tasks.
|
|
5. Never output final numbers before evidence capture.
|
|
|
|
### Strict module whitelist
|
|
|
|
Only use these exact module names in planning or reporting:
|
|
市场洞察 / 叶子类目分析 / 品牌店铺商品热榜 / 飙升品榜单 / 品牌布局分析 / 竞店布局分析 / 价格段布局 / 产品需求分析 / 需求趋势分析 / 关键词布局分析
|
|
|
|
If model generates non-whitelist names, discard and regenerate with whitelist names only.
|
|
|
|
## Analysis brief workflow (mandatory for market tasks)
|
|
|
|
When user asks for market analysis report/brief:
|
|
|
|
1. Route to yuce modules by requested metrics.
|
|
2. Collect evidence first:
|
|
- `yuce-browser open https://hq.yucekj.com`
|
|
- `yuce-browser snapshot <limit>` on at least 2 relevant views
|
|
- `yuce-browser requests --filter yucekj.com`
|
|
3. Build brief with fixed sections:
|
|
- 已抓取证据
|
|
- 核心结论
|
|
- 可执行建议
|
|
- 风险与缺口
|
|
- 下一步
|
|
4. Mark inferred statements as `推断`, never present as hard fact.
|
|
5. If evidence is incomplete, stop at gap list and ask user one precise follow-up.
|
|
|
|
### Deterministic brief command
|
|
|
|
For requests like "分析简报/结论+建议", run:
|
|
- `market-brief-report <keyword>`
|
|
|
|
Use command output as primary evidence text. Do not rewrite numeric values.
|