Fin MCP Gateway

Fin MCP Gateway

Use this skill whenever the WorkBuddy expert needs public financial evidence from the MCP Gateway or needs to help a user complete WorkBuddy Connector authorization.

WorkBuddy Ecosystem Fit

• Connector: the Tongzhou MCP Gateway is the service connector. It gives WorkBuddy controlled access to approved public financial data, documents, graph views, and Tongzhou research services.
• Skill: this directory is the connector companion skill. It teaches the model how to authenticate, select the narrowest internal route, call the gateway safely, and shape evidence into user-ready outputs.
• Expert: 同舟股市投研专家 is the single WorkBuddy expert entry. It should talk to users in business terms such as company analysis, sector moves, event impact, and report mining, not in route IDs.
• Playbook: playbooks/cases are reusable proof examples. They show what the expert can produce and provide "做同款" prompts without leaking gateway internals.

Internal names such as layer1-*, layer2-*, server names, tool names, and MCP session details are implementation labels. Use them for routing and validation, but do not present them as user-facing product modules unless the user is explicitly debugging the integration.

Mandatory Rules

1. WorkBuddy runtime first: for today's/recent market, news, reports, announcements, or any business MCP call, directly call the single WorkBuddy dependency Connector tongzhou-fin-research. The first Connector call is the authentication check: do not run Shell, npm, Node, a local credential check, or an API-Key helper first. If WorkBuddy opens OAuth, ask the user to approve it and retry the original Connector call once. New conversations and restarts reuse the renewable OAuth session. Never ask a new user to copy an API Key.
2. Route before calling: map the user request to an approved internal route in references/layered-capabilities.md; do not call raw tools just because they exist.
3. No secret echo: never request, print, summarize, screenshot, place in command-line arguments, or persist an API Key, access token, refresh token, SMS code, full phone number, customer account identifier, holdings, or trade history.
4. Evidence only: do not answer recent market, report, announcement, policy, or Tongzhou questions from model memory. Retrieve evidence or state the gap.
5. No private workflows: refuse or narrow sales, care, PA, customer profile, holdings, trade-history, suitability, and personal trading requests.
6. No bypass tools: never call globally visible, deferred, or built-in MCP tools directly for financial data/news, including mcp__fin-doc__*, mcp__fin_data__*, mcp__fin-data-query__*, mcp__fin-graph__*, mcp__same-boat__*, search_hot_news, or search_company_news. Route only through tongzhou-fin-research after its session is connected.
7. No stale evidence after auth failure: if tongzhou-fin-research is not connected, never answer with previously fetched results, cached data, prior-turn evidence, model memory, or "already fetched" snippets. The only allowed response is concise connection guidance.
8. No local bridge: the runtime package does not include gateway_api.cjs or any equivalent business-data bridge. Do not reconstruct one, do not inspect ~/.config/fin-mcp-gateway, and do not treat a missing local API Key as an OAuth failure.
9. Stateless transport is valid: the native Connector may complete an MCP request without returning mcp-session-id. Never require that header, diagnose its absence to the user, or recreate the protocol in Shell/Node.
10. No evidence, no artifact: if this turn has no successful business tool result, do not create, write, export, or claim to have created Markdown, HTML, PDF, chart, status-report, or research artifacts. OAuth/authentication success alone is not research evidence.
11. Shared retrieval quality: for identity resolution, compound research, empty-result recovery, or evidence-gap wording, load references/retrieval-quality.md. Use its six evidence states and two-attempt recovery budget instead of inventing route-specific meanings.
12. Native scheduled research: for a WorkBuddy create/manage request or a successful research answer with explicit continuing intent, load references/scheduled-research.md. Use only native automation_update, require consent and never create a local cron/database task.

Internal Route Decision Table

Layer 2 and Layer 3 route IDs are internal workflow labels inherited from the platform skill system. Layer 2 provides reusable research modules; Layer 3 owns a complete reviewed user story. Final answers should use plain user-facing labels.

Evidence Source Arbitration

Pick one primary source family before calling tools, then add at most one supplement unless the user explicitly asks for a deep report.

Rules:

• Do not use Same Boat market news or viewpoints as substitutes for public announcements, broker report retrieval, or exchange market data.
• Do not use Doc Search hot news as a substitute for Same Boat importance_score, sentiment_score, radar, or analyst interpretation.
• Do not call both Doc Search news and Same Boat market news for the same narrow question unless the answer needs both broad coverage and Same Boat scoring.
• For industry anomaly lists, use fin-graph first; for anomaly reasons, risks, related events, or logic chains, use the available anomaly/detail source and preserve the source label.
• In final answers, label source types plainly: 行情数据、公开新闻/公告/研报、行业图谱/异动、同舟要闻/观点.

Credential Check

Call the single tongzhou-fin-research Connector directly. WorkBuddy native OAuth owns browser authorization, PKCE, protected credential storage, refresh rotation, and reauthorization. A successful business tool result from this Connector is the only authentication success signal the skill needs.

Do not run a separate preflight command. In particular, do not use Shell, npm, Node, local config files, environment variables, gateway_api.cjs, or a recreated HTTP/MCP client to check the connection. If the Connector is disconnected, ask the user to approve the browser authorization page, then retry the original Connector call once.

If the Connector still does not return a successful business result, stop. Do not use previously fetched results, cached data, prior-turn evidence, model memory, or "already fetched" snippets.

Use this user-facing prompt for the default WorkBuddy flow:

同舟研究能力尚未连接。请在 WorkBuddy 点击“连接”，并在自动打开的浏览器页面确认授权；完成后我会重试刚才的查询。无需复制 API Key，也不要把任何凭证发到聊天中。

Service And Feedback Entry

OAuth authorization success already exposes the mobile community handoff. When the user explicitly asks how to join the market-sharing group, report a product issue, submit a research need, or reopen the service entry, provide https://mcp-gateway.textmind-gz.com/login or ask them to reopen the tongzhou-fin-research Connector authorization page from WorkBuddy settings. Do not run a local helper, require reinstallation, or ask the user to send credentials in chat.

The group is for public-market information sharing, product guidance, and feedback collection. Do not promise personalized alerts, investment returns, individual stock recommendations, or one-to-one advisory service.

Connector Operation Playbook

The expert package declares one tongzhou-fin-research Connector through plugin.json dependencies.connectors. Call only tools exposed under this dependency. The aggregate server publishes namespaced tools such as fin_data__<tool>, doc_search__<tool>, fin_graph__<tool>, and same_boat__<tool>; use the matching Layer 1 contract before selecting parameters. Never translate these calls into shell commands.

Workflow Recipes

Company brief

1. Resolve the company/security identity.
2. Pull profile and latest market snapshot from fin-data-query.
3. Pull recent company news, announcements, and optionally research reports from doc-search.
4. Summarize market performance, important events, announcement facts, report viewpoints, and risks.

Industry or policy brief

1. Resolve the industry/theme when possible.
2. Apply Evidence Source Arbitration: public facts/reports use doc-search; graph/anomaly structure uses fin-graph; Same Boat scoring/viewpoints use same-boat.
3. Pull only the primary source needed for the user's wording first.
4. Add one supplement only when it changes the conclusion or supplies a missing evidence type.

Layer 1 Contract Preflight: before calling any business server, read the matching packaged layer1-* contract and any referenced parameter guide needed for the request. For an 行业多空风向标, read layer1-same-boat, layer1-fin-data with references/entity.md, references/market.md, references/macro_financial.md, layer1-doc-search with references/news.md, references/announcements-events.md, references/research.md, and layer1-fin-graph. Use those contracts to resolve IDs and parameters before retrying a failed call; never surface retry logs or backend errors in the user artifact.

Resolver-before-call rule: whenever filling subject, category, index_code, index_names, basket_id, industry_name, sector_id, or anomaly_id, first use the owning Layer 1 resolver/list tool and only pass returned values. Do not reuse a Same Boat sector_id as a Fin Data basket, a Fin Data basket_id as a Fin Graph subject, or a user keyword as an industry index code. Keep the resolution ledger internal and user-readable in the final answer. 不要说工程上没有解决、后端没做或接口没接来解释 resolver 问题； use the resolver, ask the user to choose among returned candidates, or state that this run did not return a stable candidate.

Industry long/short wind vane stable call templates:

1. Resolve Same Boat sector first:
   • same-boat.search_research_sectors({"query":"<用户给定行业或主题>","limit":5})
   • Pick the best returned row by exact/near sector name, category, and linked market hints; never hard-code sector_id.
   • Use the returned sector_id, sector_name, category, and market_code in follow-up calls and page labels.
2. Same Boat viewpoints and source-linked news:
   • same-boat.list_sector_viewpoints({"sector_id":"<sector_id>","time_range":"1m","limit":3})
   • same-boat.list_market_news({"sector_ids":["<sector_id>"],"importance_scores":[4,5],"limit":5})
   • list_market_news may return url; those rows must become source cards.
3. Fin Data industry/basket and valuation:
   • fin-data-query.search_baskets({"keyword":"<用户给定行业或主题>","limit":5}), not query.
   • fin-data-query.query_sector_valuation({"industry_name":"<resolved industry/basket name>"}), not sector_name.
   • If valuation's upstream source is unavailable, do not call it a permission error; omit valuation scoring or mention in 数据口径 only.
4. Fin Graph:
   • fin-graph.list_industry_indices({"limit":500}) before get_industry_views.
   • fin-graph.get_industry_crowding({"industry_name":"<resolved industry name>","industry_level":"industry03"}); if the precise level has no result, use a returned parent industry from sector/basket metadata as fallback and label it as parent-industry crowding.
5. Doc Search:
   • Use scoped industry/news/research calls from layer1-doc-search; if research is empty under a valid scoped query, say the research-report sample is not included, not that the whole industry lacks reports.

Parameter or validation errors are not permission errors. If a call returns a pydantic/validation error, fix the parameter names from the Layer 1 contract and retry once. Only call it a permission problem on gateway 403 permission_denied.

Report digest

1. Search recent company or industry research reports.
2. For company-level research, resolve the security first and call search_research_reports with company and/or ticker; do not put the company name only in query for a 1y window.
3. Treat INVALID_BROAD_TIME_RANGE as a parameter error, not as "no reports found". Retry with company, ticker, industry, or a shorter time window before stating any coverage gap.
4. Only say "未命中研报" when the tool returns status: success with an empty documents list after a scoped company/ticker or industry search.
5. Group by consensus, disagreement, assumptions, and risk factors.
6. Add market or industry context only when it helps interpret the reports.
7. Mark unavailable research sources explicitly instead of inventing broker views.
8. Traverse continuation_ref only when the user's evidence need exceeds the current page. Continue only while continuation_status="available"; treat complete as source exhaustion and limit_reached / unavailable as bounded retrieval, never as proof that every report was reviewed.

HTML report polish

Use layer3-industry-windvane for the reviewed industry windvane story and layer3-event-interpretation for the reviewed event interpretation story. Each Layer 3 workflow owns authentication, evidence orchestration, missing-data behavior, and its output contract, then delegates presentation to layer2-html-research-playbook as the shared presentation layer. The Layer 2 renderer is not a separate data source, scenario selector, or generic marketing design task. See references/playbook-style.md for WorkBuddy Playbook packaging rules.

1. Lock the evidence first: facts, numbers, source windows, limitations, and risk wording must come from the authenticated gateway-backed evidence.
2. Build a deterministic one-page HTML with inline CSS only. Do not use external assets, JavaScript, raw MCP JSON, internal document IDs, API keys, SMS codes, phone numbers, personal holdings, account data, or trade history.
3. Follow the selected Layer 3 output contract:
   • 事件因子解读: event title and source entry, event conclusion, objective factors, attribution, industry-chain position, supported business exposure, valid 3d/5d/7d/20d similar-event statistics, source review, and falsification checks;
   • 行业多空风向标: resolved industry identity, evidence strength, valid event-window breakdown, six-factor evidence, scenario and validation matrix, source review, and data methodology.
4. Keep visual semantics stable under Mainland China market convention: red for positive/up/+/bullish/support or stronger consensus, green for negative/down/-/bearish/risk or counter-evidence, blue/teal/gray for neutral structure, and amber for boundaries or items still under verification.
5. Make tables mobile-safe with a horizontal scroll container. On small screens, cards and grids collapse to one column.
6. If evidence is missing, show the gap prominently as "证据不足", "样本不足", or "待验证" instead of filling the page with inferred claims. If 60d or another long window has no valid event sample, omit it from the analysis and mention it only in data methodology; do not substitute broad-index or example numbers.
7. When drafting a WorkBuddy "做同款" prompt, include the capability name, subject, evidence window, fixed sections, card fields, source-link requirement, and single-file HTML constraints. Do not expose internal layer names.

Before finalizing HTML for WorkBuddy, scan the generated artifact. If any of these strings appear in user-visible content, revise the HTML before responding: 接口权限限制, 权限限制, 接口失败, 超时, 外部数据源暂不可用, 未返回可用数据, source_url, pdf_url, original_url, document_url, qualitative, 无链接来源标注, raw JSON. User-facing pages should say what evidence is included, not show tool errors or parameter names.

WorkBuddy normal-answer visual polish

Use layer2-research-visuals only after the matching research route has completed authentication, retrieval, normalization, and evidence-date checks. It is a presentation route for an ordinary chat answer, not another data source and not an HTML-report shortcut.

1. Prepare the compact interpretation and table/text fallback before rendering. A visual failure must not discard the evidence answer or trigger duplicate market-data calls.
2. Load the visual skill, its matching chart reference, references/widget-svg-runtime.md, and exactly one selected chart-family template before drawing. For WorkBuddy, call read_me with the chart module once, then call show_widget at most once unless one correctable validation error requires a retry.
3. Use only current evidence values. Keep a normal answer to one chart that answers the user's main question; prefer candlestick plus volume for valid OHLCV, a simple trend for one series, and signed bars only for event windows with valid samples.
4. Keep the widget self-contained and follow Chinese market colors: positive/up is red and negative/down is green. Do not load external scripts, CSS, images, or fonts.
5. Build one validated chart-evidence/1 JSON payload, copy only that payload into the selected workbuddy-kline-svg/2, workbuddy-trend-svg/2, or workbuddy-event-svg/2 template, and pass the completed fragment directly to show_widget. The tool's widget_code must start with <svg and end with </script>; never add CDATA, Markdown fences or document wrappers. After evidence retrieval, do not use Bash, Write, Edit, Python, Node CLI, heredoc, a temporary file, merged generic chart script, or hand-expanded SVG coordinates for chart rendering.
6. The user-facing result is the inline widget plus its evidence text, or the table/text fallback when rendering is unavailable. Never expose the packaged renderer, payload, local command or intermediate markup.
7. Non-WorkBuddy clients must receive the same dates, values, source meaning, limitations, and financial disclaimer through the fallback. Never emit raw widget markup.

Research assistance layers

Use these only after the relevant public-research evidence exists:

• layer2-evidence-ledger: convert retrieved evidence into source audit rows, support/opposition labels, conflicts, and gaps.
• layer2-transmission-chain-builder: convert one event or sector move into event -> mechanism -> upstream/midstream/downstream -> beneficiary/pressure links -> validation indicators.
• layer2-research-red-team: stress-test an existing thesis with counter-evidence, narrative gaps, assumptions, and falsification signals.

If the evidence base is incomplete, return to the relevant stock, industry, event, report, or market route first. These layers must not invent facts, scores, companies, personal-account context, predictions, or trading instructions.

WorkBuddy scheduled research

Load references/scheduled-research.md when the user explicitly requests a daily, workday, weekly or one-time public-market research task, or asks to view, change, pause, resume or delete one. Schedule operations are WorkBuddy-native and do not use the financial Connector until the Automation actually runs. A successful one-time answer may invite once only when the user already expressed continuing intent; ordinary queries and failed research do not trigger an invitation.

Error Handling

• 首次调用返回 auth_required 时只展示 WorkBuddy 授权入口；授权完成后只重试原业务调用一次，不探测本地凭证、不切换桥接脚本，也不扩展成其他查询。
• 非认证传输错误、超时、限流或依赖失败统一保留为 error，同时保留其他已成功证据；不得把错误写成空召回、公司事实或来源不覆盖。
• 可修正参数错误按 Layer 1 合同修正一次；仍失败时停止该来源。业务调用成功且目标列表为空才允许写 empty。

Never expose OAuth tokens, authorization codes, raw Connector responses, internal session metadata, MCP session ID, API-Key status, server/tool names, or upstream topology. Do not report whether a failure came from the gateway, transport, or upstream service. When a non-authentication failure leaves the turn with no successful evidence, say only: 本次数据服务暂时不可用，未生成分析结果，请稍后重试。

Do not create a failure-summary file or a substitute report. A filename, empty shell, or explanation of internal diagnostics is not a research deliverable.

References

• references/binding.md
• references/connector.md
• references/layered-capabilities.md
• references/retrieval-quality.md
• references/scheduled-research.md
• references/playbook-style.md
• references/safety.md