Hanai Worth · 值见

Price has a quote; value needs research.

Hanai Worth · 值见 is a local-first A-share research workbench built on DeepSeek Harness as its Agent kernel. It connects market overview, watchlist valuation, individual stock quotes and K-line observation, master methodology analysis, open expert dialogue, report archiving, and continuous follow-up into one complete research pipeline, helping users move from "discovering a company" to "forming and continuously validating their own judgments."

DSH handles the model, Agent, tools, Sessions, streaming events, and session persistence; Hanai Worth handles securities/valuation data, watchlist groups, research materials, immutable report snapshots, and all product UI. The product has six top-level pages — "Today's Market, Watchlist & Discovery, Master Judgments, Expert Dialogue, Expert Center, Settings & Diagnostics" — plus stock, judgment, and dialogue detail pages.

The two brand trend lines cross at the evidence point: price quotes the market; research reveals value. Each judgment should trace back to evidence, method, and context.

Online Research Reports

Report	Online Page	Self-contained HTML in Repo	Content
Turning-Point Capability Audit	GitHub Pages Home	turning-point-capability-audit-2026-08-23.html	Current 10 production markers, 22 valid period cells, matched increments, candidates, and Chanlun lifecycle
A-Share One-Year Cycle Outlook	Full Gated Online Report	a-share-cycle-outlook-2026-08-25.html	Point-in-time market cross-section, Theme A–C evidence scoring, six-stock quality/valuation/relative strength, conditional historical ratios, six-gate results, and Round 4 full-state review

The cycle outlook actually invokes five expert Skills from the repo, but they are AI methodology roles sharing the same origin. The completed version first froze the 2026-08-25 market parse snapshot and official financial reports, filled in the five gates — facts, mechanism, quality, valuation, and counter-evidence — then closed the council gate via an independent council_vote in Round 4; the expert's actual pre-read inputs and the five raw outputs, prompts, and Skill hashes are stored separately and bound by SHA. The theme-mapping cross-check and the stock-gate cross-check are not independent votes; specific model and reasoning configurations were not frozen in the artifacts; and the report does not provide uncalibrated probabilities, target prices, or return promises.

UI Preview

Dark mode keeps the original client's density, page layout, and A-share red-up/green-down semantics; the market heatmap is drawn with ECharts treemap, area mapped to sector turnover.

Light mode only swaps semantic color tokens; sidebar, top bar, cards, tables, and charts keep their positions and sizes.

Stock detail shows quotes and value side by side: left has K-lines, snapshot, fundamentals; right has value judgment and curve; gold = Master Value, blue = price, red/green = over/undervalued.

Turning-point markers place recalculable volume-price conditions on daily/weekly/monthly K-lines. Below: China Ping An hit 低位破低反包 on 2026-07-01; the hover card shows trigger conditions, 20-day direction frequency, sample count, same-scenario comparison and research cutoff date; the toolbar toggles all markers with one click.

Percentages in the screenshot are historical conditional frequencies as of 2026-08-20, not a forecast of Ping An's next move, and do not constitute investment advice.

From Quotes to Research: The Full Chain

1. See the market: Six major indices, market breadth, sector/concept heatmaps, and gain/loss/turnover/turnover-rate leaderboards together present the day's structure.
2. Build watchlists: Search stocks, add, move, and manage symbols by group; refresh the current group and compare quotes, valuation, and since-added performance.
3. Read a stock: Combine forward-adjusted K-lines, moving averages, volume-price observation markers, quote snapshot, fundamentals, fair value, and value curves to build context.
4. Open dialogue: Skip stock selection and pick an expert directly, discussing industry cycles, business models, recent events, or decision questions in depth.
5. Launch a judgment: Choose a stock and a master; the Agent independently retrieves and verifies public materials to produce an archivable full Markdown report.
6. Keep verifying: Open dialogues and judgment follow-ups each reuse their original DSH Session; judgments can also review tool steps or explicitly revise the report.

Capabilities

Today's Market & Watchlist

• Market Overview: Six major indices, limit-up/down and advance/decline counts, combined turnover, industry/concept treemap and four rankings load in parallel; heatmap area = turnover, colors follow A-share red-up/green-down.
• Watchlist Groups: Default groups always retained; supports creating, renaming, deleting custom groups, and adding, moving, removing stocks with tri-state sorting.
• Watchlist Quotes: Latest price, change %, turnover, turnover rate, market cap, PE, PB, add date, performance since added; quote status distinguishes fresh, stale, unavailable.
• Watchlist Valuation: Fair value loaded async per group via Value Master Network with concurrency limits, not blocking the quote table; shows absolute and percentage gap to current price. Red for upside, green for downside.
• Loading & Refresh: Manual refresh of current group; quotes and valuation show loading/failed/no-data states separately; single stock failure won't drag down the group.

Stock Quotes, Valuation & K-line

• Quote & Financial Snapshot: Latest quote, OHLC, volume/turnover, turnover rate, volume ratio, main net inflow, market cap, PE, PB, ROE, EPS, revenue, profit margin, debt-to-asset ratio.
• Multi-period Quotes: Intraday, daily, weekly, monthly K-lines; all use forward-adjusted prices. Daily K loads on demand first screen, drag left for more; weekly/monthly provide full history. Currently viewed K refreshes every 15s, same-date overwrites, new periods append; switching back triggers immediate refresh.
• Dual MA Mode: Toggle between short-term MA5 / MA10 and mid-term MA20 / MA60; MAs computed on close of current period.
• Turning-point Markers: Daily/weekly/monthly K retain "massive-volume divergence, weak close, deep-drop volume, strong close, long shadow, rebound stabilization"; daily K adds "low-level breakdown reversal engulf, golden-needle watch, golden-needle confirm, high-level long upper shadow". Same-period markers can stack; hover shows quote, MAs, trigger conditions and historical evidence in one tooltip.
• Dynamic Recalculation & Toggles: Recalculate only when latest K's OHLCV/turnover changes; latest bar shows "dynamic calculation", historical bars show "close confirmed". Toggle on by default; off hides markers/evidence only, doesn't stop refreshes or affect K-lines, volume, MAs.
• Value Judgment: Fair value loading uses independent skeleton/animation; placeholder only when no data; valuation summary, five-dim radar, price/Master Value curve are independent, not blocked by source failure.
• Traceable Sources: Charts/cards retain source, fetch time, latency/cache status; missing fields show — or hidden, never 0.

Turning-point markers are volume-price observation hints filtered by historical condition frequency, not buy/sell points or return guarantees. The original six categories show independent evidence on daily/weekly/monthly K-lines; the new 低位破低反包, golden-needle watch/confirm and 高位巨量长上影 have daily-only evidence, so are not extrapolated to weekly/monthly.

Full turning-point capability audit: view the GitHub Pages online report or the self-contained HTML in the repo. The report reconciles, one by one, the current 10 production markers, 22 valid period cells, 16 frozen candidates, matched increments, confidence intervals and planned shadow-validation directions.

How Markers Are Calculated

Live triggers don't use an LLM, nor read news, fundamentals, or future data; they compute deterministically on fetched forward-adjusted OHLCV. Same K-line input always yields same markers. Fewer than 121 bars of the current period means no marker shown, ensuring MAs, drawdowns, and volatility are warmed up.

Metric	Basis
Trend	MA5 / MA10 / MA20 / MA60, returns over the prior 10/20 periods, drawdown from the prior 60-period high
Volume	Volume / VMA20; the new daily V0 uses VMA20⁻ that excludes the current day, so today's outsized volume cannot inflate its own baseline
Volatility	TR and ATR20⁻ that excludes the current day
Candle position	CLV=(Close-Low)/(High-Low), body, upper wick, and lower wick as proportions of the period's total range
Structure	Prior high/prior low, 20/60-period highs and lows, and the number of periods each has stayed below its own MA20
Confirmation	Raw patterns are marked on the anchor day; patterns that require confirmation are marked only on the actual confirmation day — future confirmations are never backfilled into the past
Deduplication	Same stock and same pattern are cooled down within the pre-defined observation window to avoid flagging consecutive, similar candles repeatedly

产品中的十类标记分为三种语义：

Product semantic	Marker	Timeframe	Meaning
Volume-price observation	分 / 弱 / 深 / 强 / 影 / 稳	Daily, weekly, monthly	Describes outsized volume divergence after a rally, or volume expansion, strong closes, long wicks, and stabilization in a deep-drawdown zone
Low-zone observation/confirmation	包 / 针 / 确	Daily K only	包 is a bearish engulfing that breaks the prior low then reclaims the prior high; 针 is only a golden-needle anchor; 确 appears only when a true breakout of the anchor high and MA5 occurs within the next 1–3 days
Risk observation	拒	Daily K only	A new high within an uptrend accompanied by outsized volume, a long upper wick, and a weak close; flags potential distribution stalling, not a mechanical sell signal

精确阈值、显示冷却与 tooltip 口径见变盘点产品决策。

全市场回测与验证

FULL_A_TURNING_V0 研究冻结了 16 个原始/确认规则，先定规则和主观察窗口，再统计结果，避免看到漂亮数字后临时改参数。研究请求 5,813 只当前及历史 A 股证券，成功取得 5,809 只行情；其中 5,704 只达到至少 120 根预热要求，共分析 9,813,872 根有效可交易日 K，集合日期范围为 2014-01-02 至 2026-08-20。公开行情来自新浪、腾讯和东方财富的复权日线；北交所退市历史、逐日 ST 状态和完整退市收益仍不完备，因此这里称“实质性全市场扩展”，不称无偏的全历史 A 股。

统一回测时钟与交易约束：

• 信号只使用 t 收盘时已知的 OHLCV，滚动量、ATR、前高和前低使用 t 之前的数据；
• 最早执行价为信号后下一根可交易 K 的开盘，H 日在 signal + H 收盘结算；零成交/停牌行不作为可交易 K；
• 买侧正收益频率扣除约 0.312% 固定往返摩擦；风险侧只统计未来原始股价下跌频率，不伪装成可做空收益；
• 信号日前 20 根成交额中位数至少 3,000 万元，5,000 万元作为预先声明的流动性敏感性；
• 每只股票每个信号阶段按主窗口去重，年度和时代分段要求退出日也留在分段内，未成熟事件直接删除而不是尾部强平。

下表是扩展未见股票的主窗口结果；“方向频率”是回顾性的条件比例，俗称“胜率”但不是下一次概率。买侧平均收益已扣除固定摩擦，风险侧平均方向收益表示平均下跌幅度。

冻结形态	主窗口与方向	样本	历史方向频率	平均方向收益	同场景对照	产品结论
低位破低反包	20 日上涨	3,581	65.60%	+7.12%	65.75%	显示 包；形态独立增量未证实
深跌巨量强收	20 日上涨	1,770	64.18%	+9.68%	66.21%	保留相邻旧观察；巨量增量未证实
金针探底观察	20 日上涨	18,937	50.85%	+1.74%	50.34%	显示蓝色研究锚点，不称买点
金针突破确认	20 日上涨	10,650	51.81%	+1.53%	50.80%	确认未提升为高胜率，只显示 确
高位巨量长上影	5 日下跌	5,721	56.58%	+0.15%	55.46%	显示 拒风险观察，不称卖点
巨量后缩量不破转强	20 日上涨	75	50.67%	+2.07%	51.75%	样本过少，不进入默认标记
极缩量后脱离平台	20 日上涨	22,495	47.68%	+0.84%	49.53%	明确否决为买点

验证设计不仅看原始方向频率：

1. 旧研究涉及的 249 只股票整体放入 prior_studied，不参与扩展未见股票主结果；其余股票按代码 SHA-256 固定拆分为开发 50%、验证 30%、测试 20%。
2. 每个事件匹配同一交易日、同方向、同板块、相近前置涨跌幅和流动性层的普通股票，判断形态是否优于当时共同市场环境。
3. 方向与收益使用“股票 + 信号月份”双向聚类区间；16 个主假设的匹配方向增量使用 Holm 5% 家族错误率校正。
4. 额外检查时代/年度、主板/创业板/科创板/北交所、3,000万/5,000万元流动性、去掉最佳 5% 事件和更高摩擦后的稳定性。
5. 严格上线门槛还要求非重叠样本至少 200、原始方向率至少 60%、聚类区间下界高于 50%、匹配方向及收益增量区间下界均高于 0、跨年度和验证/测试股票折稳定。本轮通过数为 0/16，因此产品没有新增任何“已验证高胜率买点”。

截至 2026-08-20 的数据已经参与规则研究，只能作为回顾性证据。从 2026-08-21 起应按冻结 V0 规则持续记录前瞻影子事件，积累足够样本后再做一次不改参数的检验。完整覆盖、分层、聚类区间、匹配对照、否证结果和已知偏差见全市场 A 股日 K 变盘点研究，机器可读结果见冻结 JSON，复现脚本见full-market-turning-point-study.py。完整研究索引见设计文档。

专家开放对谈

• 不必先选股票：可直接讨论行业周期、商业模式、市场情绪、决策困境或近期事件，不生成 REPORT.md。
• 持久开放会话：每次对谈绑定独立专家快照与 DSH Session，支持正常追问、检索、工具调用、队列、steer、取消、审批和历史恢复。
• 专用对谈工作台：左侧保留对谈历史，右侧消息区占满剩余高度；支持深链接和刷新恢复，思考与工具过程默认折叠，输入区固定在当前会话底部。
• 六位对谈专家：段永平、混江龙、查理·芒格、沃伦·巴菲特、散户乙均支持研判与开放对谈；孙宇晨视角仅进入开放对谈。
• 散户乙视角：按 nuwa 的本地语料流程，从《从冲浪到潜水》发言汇编提炼股权思维、可持续 ROE、分红复利和底线回报；保留 PDF 页码、观点演化和汇编归属边界。详见能力包与蒸馏记录。
• 孙宇晨视角边界：参考用户指定的开源能力包，用于分析行业周期、注意力迁移和叙事竞争；创建对谈时明确提示 AI 模拟边界，首次回答完成身份披露，详情页不重复占用消息空间；具体时效事实必须先检索核验。
• 周期可证伪：“永远缺某种资源”只作为假设，从需求、供给、库存与利用率、资本开支、价格利润和拥挤度检查，并给出反证与失效条件。
• 单一事实源：Hanai SQLite 只保存标题、专家和 opaque dshSessionId；消息、工具与 Turn 历史仍只在 DSH。

大师研判与持续对话

• 单专家独立研判：支持段永平、查理·芒格、沃伦·巴菲特、混江龙和散户乙五套方法论；每次研判绑定独立工作区和持久 DSH Session。
• 可核验报告：保留 preparing → running → verifying → completed/failed 状态、实时执行过程、失败原因、归档信息、不可变报告版本、哈希与文件大小。
• 报告默认、对话延续：完成后默认打开研判报告，也可切换到“继续对话”，沿用原 dshSessionId 追问；普通追问不会静默创建新报告版本。
• Markdown 与过程展示：报告和对话正确渲染标题、列表、表格、引用、链接、行内代码和代码块；思考与工具活动按轮次紧凑折叠，详细参数和结果按需展开。
• 运行中交互：支持排队发送、立即插话、编辑/移除队列消息、取消运行，以及工具批准和结构化问题回复。
• 安全删除：仅已完成或失败且会话不在运行的研判允许删除；二次确认后移除全部本地报告/工作文件并归档对应 DSH Session，进行中的研判受到保护。

工作台与运行管理

• 统一交互系统：按钮、输入框、选择器、焦点、禁用和危险操作使用一致的语义色与明暗主题 token，并保留 A 股红涨绿跌业务色。
• 亮色/黑夜模式：两套主题只替换语义色彩，不改变页面结构、图表数据或业务含义。
• 全局搜索与深链接：可按代码、名称或拼音搜索股票；#/dashboard、#/watch、#/judgements、#/expert-chats、#/personas、#/settings 及详情页支持刷新、前进、后退和直接打开。
• 设置与诊断：在工作台内管理 DSH Credentials、默认模型和主题，查看 Agent、数据源、缓存、本地存储与版本状态，并可打开数据目录或清理行情/估值缓存。
• 完全自有界面：使用 React 18、DSH Slot/Runtime、ECharts 和 CSS Modules；不显示或复用 DSH 原生聊天 UI。

架构

flowchart LR
  U["用户"] --> W["Hanai React Workbench\nshell.overlay"]
  W --> R["/hanai Connection RPC"]
  W --> C["DSH Credentials / Models"]
  W --> S["DSH Session Runtime"]
  R --> H["Hanai Host Service"]
  H --> D["SQLite / Reports / Masters"]
  H --> M["Eastmoney / Tencent / GuruFocus"]
  H --> S
  S --> A["DSH Agent + Tools"]
  A --> L["DeepSeek Models"]
  S --> P["DSH Session Persistence"]

关键边界：

• DSH 是聊天与 Agent 的唯一事实源；Hanai 不建立第二套 messages 或 turns 表。
• Hanai SQLite 只保存自选、证券主数据、研判/专家对谈索引、报告版本和 opaque dshSessionId。
• 报告是 Hanai 封存的业务快照；工作区 REPORT.md 只是 Agent 可写的生成副本。
• 浏览器不接触文件系统或 SQLite；全部业务写入经由同源 /hanai RPC。

更完整的设计见 总体架构 与 ADR 索引。

运行要求

• Node.js ^22.19.0 或 >=24.0.0
• pnpm 11.7.0
• DeepSeek Harness 0.1.1-rc.2；DSH 仍处于 pre-release，升级到其它 rc 前必须重新验证，CLI、Web App 与 Hanai 应使用同一版本
• 一个 DeepSeek API Key（只在实际运行 Agent 时需要）

DSH 仍处于 pre-release，rc 之间不承诺兼容。仓库把 Host、Client 和 profile 装配都纳入兼容性检查，但升级前仍应运行完整门禁。

从源码安装

git clone git@github.com:hanai-labs/worth-dsh.git
cd worth-dsh
pnpm install
pnpm run build
pnpm run profile:install -- --package .
pnpm run profile:verify
dsh --profile hanai-investment

安装器会创建或安全迁移独立的 hanai-investment Profile。最终 Bundle 顺序固定为 DSH Base、DSH Web App、Hanai；只有 hanai-investment-dsh 是 Profile dependency。Base 与 Web App 必须由当前 DSH CLI 的 installation fallback 提供，不能再用 dsh plugin add @deepseek-ai/dsh-web-app 安装到 Profile，否则相同版本的 DSH runtime 仍可能被加载成两个模块实例。安装器会拒绝修改 web、headless 等保留 Profile，也会在目标 Profile 含无关依赖或 Bundle 时停止。

通用 DSH Web 仍按原方式启动：

dsh web

两者可以使用不同端口同时运行。详见 ADR-0003。

安装发布包或修复旧 Profile

在包含本仓库安装脚本的发布目录中，把 --package 换成 npm 包名即可。重复执行是安全的，也会迁移早期错误安装过 Web App dependency 的 Profile：

pnpm run build
pnpm run profile:install -- --package hanai-investment-dsh
pnpm run profile:verify
dsh --profile hanai-investment

迁移前请先停止正在运行的 dsh --profile hanai-investment 进程。不要手工执行 dsh plugin ... add @deepseek-ai/dsh-web-app；它会重新引入 Profile-local DSH runtime shadow。

数据与隐私

Hanai 业务数据默认写入：

~/.hanai-investment-dsh/
├── db/hanai.sqlite
├── cache/
├── judgements/<id>/workspace/
├── judgements/<id>/reports/<version>/
└── expert-chats/<id>/workspace/

新版不会检测、读取、导入、修改或删除旧版数据目录。首次启动会建立一套空数据库，自选和研判需要重新创建。

以下内容仍由当前 $DSH_HOME 管理：

• DeepSeek Key 与模型设置；
• Session 事件、消息、工具历史；
• 聊天附件和 Profile 安装状态。

Hanai 数据根默认权限为 0700，普通数据文件为 0600。API Key 是 write-only secret：页面提交后清空输入，RPC 不返回明文，日志和报告也不得包含它。

行情与来源语义

数据源必须把“真实值”和“可用性”一起交给 UI：

• 东方财富实时集群可用时标记 fresh；
• Node TLS 环境被实时集群拒绝时，非历史行情降级到东方财富延迟源；
• 分时和前复权 K 线在必要时降级到腾讯行情；日 K 支持向左拖动分段补齐历史，周/月 K 返回完整历史；
• 自选行情先返回，价值大师网合理估值再按组异步补齐并按日缓存；两条链路互不阻塞；
• 估值加载中、供应商无数据和请求失败是三种不同 UI 状态；合理估值或成交额缺失时不渲染虚假值；
• 最近成功快照标记 stale，完全不可用则显示 unavailable；
• 缺失值始终显示为 —，绝不解释成 0；
• 页面不合成不存在的指数走势，也不把延迟或缓存数据标成 LIVE。

GuruFocus 接口仅作为个人研究原型使用，遵循页面声明的来源、缓存时间与再分发限制。生产或商业部署应替换为有正式授权和 SLA 的数据供应商。

开发与验证

pnpm run typecheck
pnpm run test
pnpm run build
pnpm run pack:check

一键执行全部门禁：

pnpm run check

门禁覆盖：

• Provider 解析、降级、缓存和证券同步；
• SQLite migration、事务、权限和数据隔离；
• 报告校验、修复、原子封存、哈希与版本；
• 研判删除约束、Session 归档与本地文件清理；
• DSH Session 报告、开放对谈与普通追问生命周期；
• 自绘聊天的 Markdown、紧凑过程、pending、queue/steer、双语境文案和固定高度布局；
• 前复权历史加载、日/周/月最新 K 的 15 秒刷新与同日期合并、MA5/10 与 MA20/60、十类变盘点、动态/历史状态、显示开关及 tooltip；
• 自选合理估值的异步批量加载、缓存、失败隔离和距现价计算；
• DSH Client ModuleLoader 单文件协议；
• npm allowlist、入口、source map、大师资产与旧数据路径硬隔离。

真实装配验证使用临时 DSH_HOME 安装 hanai-investment Profile，再以随机 loopback 端口启动 Host/Web；不会触碰用户的官方 web Profile。

完整的已验证启动步骤、首次设置、浏览器矩阵和故障排查见 启动与验收报告。逐页功能与布局约束见 客户端迁移与验收基线。

目录结构

packages/
├── contracts/          JSON-safe Host/Client 合约
├── domain/             SQLite、报告、专家对谈、行情、估值、证券与自选领域逻辑
├── host/               Cordis Host、/hanai RPC、DSH Session 编排
├── client-workbench/   全屏 React 产品工作台
├── client-chat/        Hanai 自绘 DSH Session 对话
└── masters/            六位专家的 Skill、能力分流与参考资料
tooling/
└── dsh-client-bundle/  树外 DSH Client closure 构建适配器
scripts/
├── research/           可复现的量价与 K 线条件研究脚本
└── ...                 profile 安装、校验与发布门禁
docs/
├── research-data/      冻结口径、样本清单与可复现研究结果
└── ...                 架构、ADR、产品决策与验收文档

当前边界

• shell.overlay 是 DSH AppFrame 的全屏插件画布，不是弹窗、iframe 或命令行 Shell。它让 Hanai 在不 fork DSH 的前提下拥有完整页面；Workbench 通过 location.hash 提供页面深链接，不要求 DSH 增加通用 Router Slot。
• DSH 尚未发布稳定的树外 Client Plugin 构建 SDK，因此仓库维护了一个最小、版本锁定的 bundler adapter。
• DSH Session 删除与跨版本迁移能力仍有限；备份持续对话时，需要同时保留 Hanai 数据根和 DSH Session 数据。
• 本项目是研究辅助工具，不构成投资建议，不承诺数据实时性、完整性或投资收益。

License

MIT

客户端 bundle 内联依赖的许可证与声明见 THIRD_PARTY_NOTICES.md。