chart

name

chart

description

Render financial charts via the local chart web app (workspace at the repo root — Hono server + React front end, port 1792). Four chart types: intraday capital-flow line (`flow`) and cross-symbol signed-bar comparison (`cohort`) — both Recharts — plus SEPA strategy dashboard (`sepa`) and short-term multi-timeframe prediction dashboard (`intraday`) — both TradingView Lightweight Charts. Multi-timeframe K-line review lives inside `intraday` (the standalone kline chart type was removed). The server fetches Longbridge data itself (kline / capital flow) and computes all indicators (MA, MACD, RS, trend template, volume profile, divergence/beichi detection) in TypeScript; the caller only POSTs `{type, symbol, ...}` to `/api/charts` and gets back `{id, url, technicals?}`. Charts persist as data JSON under `journal/charts/data/`. `url` now points at where the chart actually lives: sepa/intraday land on the symbol page (`http://localhost:1792/symbol/<SYM>?analysis=<id>`), flow/cohort land on the home page for that date (`http://localhost:1792/?date=YYYY-MM-DD`); old `http://localhost:1792/charts/<id>` links still work and redirect there. Triggers: 出图、生成图表、画 K 线、画资金流曲线、画对比图、SEPA 仪表盘、 短线预测、多周期K线、MACD、入场判断可视化、可视化、render chart, plot, visualise, sepa dashboard, intraday prediction dashboard.

chart

Creates charts through the local chart app so the user can browse them in one place instead of squinting at tables. The server pulls market data and computes everything; charts are stored as versioned data JSON and always rendered by the latest front-end code.

Response language: match the user — 简体 / 繁體 / English.

When to call

• After running longbridge capital --flow context or when the user wants a flow visual ⇒ flow
• For K-line review (multi-timeframe candles + MACD + auto signals) ⇒ intraday
• After collecting cumulative net inflow across a cohort of symbols ⇒ cohort
• After running sepa-strategy on a single name ⇒ sepa
• When inside intraday-signal ⇒ intraday (two-call pattern: POST preview → PATCH prediction)
• When inside capital-rotation / market-session-tracker / stock-deep-dive, call this as the LAST step and append the chart URL to the markdown journal entry.

Skip when the user only wants a single number or a tiny series — a Unicode sparkline in the chat reply is faster.

Data-root scenario (kansoku-cli)

在 Kansoku 数据目录里跑（外部 Claude Code、无本地 dev server），走 kansoku-cli 而不是 HTTP。判据：环境变量 $KANSOKU_CLI 有值 → 走 CLI； 无 → 回退 HTTP。

CLI 与 HTTP 契约一致——body 就是原本 POST /api/charts 的 body， 返回 {id, url, technicals?} 也不变；CLI 侧额外多带一个 deepLink 字段（见下方数据目录场景说明）。

先加载 runtime.env（这一步只需做一次）：

set -a
[ -f .kansoku-agent-kit/runtime.env ] && . .kansoku-agent-kit/runtime.env
set +a

之后：

HTTP 版	CLI 等价
curl -s http://localhost:1792/api/health	"$KANSOKU_CLI" info kit-version
POST /api/charts body 为 X	printf '%s' "$X" | "$KANSOKU_CLI" chart create --type <T> --symbol <S> --json-input -
GET /api/charts?type=&symbol=	"$KANSOKU_CLI" chart list [--symbol X]
GET /api/charts/<id>	"$KANSOKU_CLI" chart get <id>
PATCH /api/charts/<id>	(暂不支持——PATCH 仍走 HTTP；数据目录场景 PATCH 目前只在 intraday-signal 里出现，届时 fallback HTTP)
打开 data.url（浏览器访问 http://localhost:1792/...）	打开 data.deepLink（kansoku://route/...，见下）

body/response 完全一致，写入的图表 JSON 落在 journal/charts/data/<id>.json （跟 HTTP 版本相同）。

数据目录场景通常没有本机 dev server（Kit 用户不会跑 pnpm start）， data.url 里的 http://localhost:1792/... 打不开。CLI 的 chart create 响应现在多带一个 deepLink 字段（url 照旧保留，向后兼容 dev 场景）， 指向 kansoku://route/... 这个自定义协议。想让用户看到图表时，打开这个 deep link 即可——它会启动 Kansoku.app（若已在运行则直接切到前台）并跳转到 对应图表页面：

open "$(echo "$response" | jq -r .deepLink)"

macOS 用 open，Windows 用 start，Linux 用 xdg-open——三者都认协议处理器， 换平台只需换这一条命令。

Server lifecycle

The app must be running before any API call:

curl -s http://localhost:1792/api/health # {"ok":true,...} = up

CLI 等价："$KANSOKU_CLI" info kit-version

If it is down, start it (long-running process — use run_in_background):

pnpm start # serves API + built web UI on :1792

First-time setup only: pnpm install && pnpm build.

API

Base URL http://localhost:1792. All responses follow the {ok, data, meta} / {ok:false, error, hint} contract.

Endpoint	Purpose
GET /api/health	liveness check
GET /api/charts?type=&symbol=&limit=	list chart metas (newest first)
POST /api/charts	create a chart; body below
GET /api/charts/:id	full chart doc
GET /api/charts/:id/built?count=	ephemeral intraday rebuild with a larger bar window (history view; max 1000, never persisted)
PATCH /api/charts/:id	merge fields into input and rebuild (e.g. add prediction)
DELETE /api/charts/:id	remove a chart
GET /api/legacy	list old single-file HTML archives (served at /legacy/<file>)
GET /api/stream/quotes?extra=	SSE quote snapshots (watchlist ∪ positions ∪ extra), 10s cadence
GET /api/stream/charts/:id	SSE live rebuilds for flow/intraday charts, 60s cadence
GET /api/symbols/:sym/{flow,benchmark,position,analyses,latest}	live per-symbol cockpit data (server-computed, never AI)

The stream endpoints power the web UI's realtime display; the AI workflow never needs them — created charts update themselves in the browser while open, and the persisted JSON stays frozen at analysis time.

Symbol cockpit (/symbol/<SYM>)

Every symbol also gets a stable dashboard URL, http://localhost:1792/symbol/<SYM>, that aggregates live data with the symbol's latest intraday/sepa analysis — it is the caller-facing counterpart of GET /api/symbols/:sym/*. The same page also serves the frozen per-analysis view: appending ?analysis=<id> pins it to one specific chart doc (this is what data.url returns for sepa/intraday charts) instead of following the latest one live; a banner offers to jump back to latest if a newer analysis has since landed. Old /charts/<id> links still work — they resolve the chart's type client-side and redirect to this page (or to the home page for flow/cohort charts). During regular session the cockpit also carries a live AI comment stream and auto-reassessment (commentator + escalated analyst, gated on the comment/analyst model settings configured at /settings) whose output shares the manual intraday-signal format. These /api/symbols/* routes are server plumbing for that page — direct callers rarely need them, listed here only for completeness:

• GET /api/symbols/:sym/flow — today's capital-flow curve + large/medium/small distribution (live, via longbridge capital)
• GET /api/symbols/:sym/benchmark — SMH/QQQ normalized same-session comparison (live, via longbridge kline)
• GET /api/symbols/:sym/position — shares/cost/unrealized + distance to stop/target from the latest analysis's entry plan (live)
• GET /api/symbols/:sym/analyses — past intraday analyses for this symbol with mechanical outcome judgments (hit_target / hit_stop / open, computed server-side from post-anchor bars — never AI recall)
• GET /api/symbols/:sym/latest — the latest intraday chart doc in full, plus prediction_stale

The client-side indicator toggle bar (show/hide 金叉死叉、自动背离、自动背驰、123 结构、SB 结构、K线形态、AI 标注、价位线、EMA 均线; state in localStorage) has no API surface — it's a pure front-end feature on both the cockpit and archive pages. Swing 高低点 are not one of the toggles — they're baked into the divergence/beichi/pattern123/SB detectors as the underlying pivot data, not a standalone overlay (deliberate deviation from the earlier spec draft). SB（第二次 突破/跌破，对应 Al Brooks 的 High 2 / Low 2）标注顺势回调后第二次尝试冲破前高/前低 的结构：图上灰色 H1/L1 是失败的第一次尝试，金色 H2/L2 箭头是确认成功的第二次突破； technicals 里对应 second_breakouts 这一行。

POST body per type

The server fetches Longbridge data itself when symbol is given; pass data (or kline / timeframes) only to override with hand-assembled rows.

// flow — server runs `longbridge capital <SYM> --flow`
{ "type": "flow", "symbol": "MU.US", "subtitle": "单位推断为千 USD · 仅供参考" }

// cohort — data is always caller-assembled
{ "type": "cohort", "title": "存储 vs Mag 7 主力净流",
  "data": [{ "symbol": "MU", "value": -17087 }, { "symbol": "NVDA", "value": 9540 }] }

// sepa — server pulls 260 day bars + SPY.US automatically
{ "type": "sepa", "symbol": "MRVL.US"  
               
      



      
                             
               
                                    

CLI 等价（body 走 stdin，--json-input -）：

• flow → "$KANSOKU_CLI" chart create --type flow --symbol MU.US --json-input -
• cohort → "$KANSOKU_CLI" chart create --type cohort --json-input -（无单一 symbol，省略 --symbol）
• sepa → "$KANSOKU_CLI" chart create --type sepa --symbol MRVL.US --json-input -
• intraday → "$KANSOKU_CLI" chart create --type intraday --symbol MU.US --json-input -

Success returns data.id, data.url (paste this into journal entries), plus type-specific meta: sepa → verdict_tier / passes / fails / bars; intraday → mode / bars / technicals.

sepa context schema

All fields optional:

{
  "earnings_dates": ["2026-05-29"], // E markers on those bars
  "stage": "Stage 2 末期", // 阶段判断 sidebar card
  "stage_note": "Stage 3 顶部嫌疑",
  "base_count": "3-4 (减半仓)",
  "pattern": "无可买（扩张振幅）",
  "verdict": {
    // override the auto verdict
    "tier": "watch", // pass / watch / buy
    "label": "👀 WATCH LIST",
    "color": "#ffc107",
    "reason": "...",
  },
  "entry_plan": {
    // 入场计划 card + price lines
    "pivot": 260.0, // required: consolidation-range high
      
      
      
     
      
  
   
    
    
       
       
        
       
       
       
    
  
    
        

Derived values (auto-computed server-side): buy_zone_high = pivot × 1.05, target1/2 = pivot × (1 + pct/100), R/R = (target2 − pivot) / (pivot − stop) — based on T2, not T1, because T1 is the SEPA Phase-2 partial exit. R/R < 2:1 renders a red warning.

Verdict auto-detection (when context.verdict omitted): any trend-template fail → PASS 🚫; all 8 pass + price ≥ 25% above MA50 → WATCH · Extended 👀; all 8 pass otherwise → WATCH · No pattern detected 👀. STRONG BUY ✅ is never auto-emitted — pass context.verdict after manually confirming a valid pattern

• pivot ±5% buy zone.

Auto markers on the main K-line: earnings (context.earnings_dates), climax top (volume ≥ 2.5×20MA + red close + local high), MA50/MA200 breakdowns, 52w high. Hardcoded lines: 52w high/low, MA50 × 1.25 extended warning.

intraday two-call pattern

1. POST without prediction → preview. Read data.technicals from the response: per timeframe last_dif / last_dea / last_hist, last_vwap (当日 session VWAP，m5/m15), emas (latest fast/mid/slow EMA values — price vs EMA stack tells the short-term trend posture), recent swing highs/lows, last_cross (金叉/死叉), divergence_candidates, beichi_candidates, second_breakouts (SB 结构， H2/L2). Also read meta.day_context: daily_trend (up/down/range vs 日线 MA20/MA50), daily_ma20/daily_ma50, high_20d/low_20d, prev_day (昨日高/低/收), pre_market 区间, opening_range (开盘前 30 分钟), vwap — the server draws these as the VWAP line and 日内参照位 price lines on the chart. Read these numbers — don't eyeball candles.
2. PATCH /api/charts/:id with {"prediction": {...}} → final dashboard. Add "refresh": true to any PATCH to refetch the latest bars (incl. pre/post market) and recompute everything before rebuilding — same id, same URL.

CLI 等价：第 1 步（POST 预览）可用 "$KANSOKU_CLI" chart create --type intraday --symbol MU.US --json-input -； 第 2 步（PATCH 写入 prediction）目前无 CLI 等价，仍走 HTTP。

prediction schema:

{
  "direction": "short", // long | short | neutral
  "anchor": { "timeframe": "m15", "time": "2026-07-01T17:00:00Z", "price": 1049.81 },
  "scenarios": [
    // ≥ 2, probabilities ≈ 100
    { "label": "继续探底", "probability": 45, "path": "...", "trigger": "..." },
  ],
  "range_bound_plan": {
    "condition": "...",
    "long_tactic": "...",
    "short_tactic" 
     
     
  
  
   
     
     
     
     
     
   
   
    
    
       
       
       
       
       
        
    
  
   
    
    
       
       
       
       
       
       
    
  

R/R is direction-aware (long: risk = entry−stop; short: risk = stop−entry); the sidebar flags rr < 2:1 in red. entry_plan can carry structured level context so the chart explains why a point was selected instead of hiding the reason in prose:

{
  "entry": 61.1,
  "stop": 62.52,
  "target1": 60.0, // optional explicit target price; overrides pct-derived price
  "target2": 57.92,
  "rationale": "反弹到 60.90-61.35 压力带后受阻才入场。",
  "stop_note": "站回上一段反弹高点，空头计划失效。",
  "entry_zone": { "kind": "resistance", "label": "反弹压力带", "low": 60.9, "high": 61.35 },
  "target1_label": "T1 · 日内低点",
  "target1_note": "整数位和日内低点，首次触及先看是否止跌。",
  "target1_zone": {         
   
   
            

entry_plan.entry_zone and target1_zone / target2_zone are explanation context for the right panel only; they are not drawn as chart zones and should not be named 入场区, T1 区域, or T2 区域. Put only genuine upper supply areas in top-level price_zones with kind: "resistance"; those are rendered as chart boundaries and in the sidebar's key-zone section.

Supported zone kinds remain entry, stop, target, support, resistance, invalidation, watch, but the intraday chart-zone overlay intentionally filters to explicit resistance zones. MACD structure signals + simplified 背离/背驰 are auto-detected and drawn on every render regardless of prediction. Every DIF/DEA cross is classified by zero-line position（零上/零下金叉、零上/零下死叉） plus structural patterns: 二次金叉（零下双金叉且低点抬高 → 底部确认）、空中加油 （零上二次金叉 → 强势延续）、二次死叉（顶部确认 / 空头中继）、上穿/下穿零轴 （趋势确认，last 2 bars marked 待确认 with a ?）. Each marker's tooltip carries the implication text; technicals.<tf>.structure_signals (last 6) and zero_tangle（DIF 贴零轴缠绕 = 震荡市，交叉信号失效）expose the same data to the analysis workflow — read them in the preview step. The swing-based divergence auto-detectors only fire on confirmed swing pivots — the last 1-2 bars can never be flagged; read last_dif/last_dea/last_hist directly for the newest bar.

14 classic K-line patterns are also auto-detected on the main pane (arrow markers, tooltip carries the implication): 单根——锤子线/上吊线/倒锤子/射击之星; 双根——看涨吞没/看跌吞没/乌云盖顶/刺透形态/看涨孕线/看跌孕线; 三根——启明星/ 黄昏星/红三兵/三只乌鸦. All require a trend-context filter (4-bar net move into the pattern) plus a body-size filter vs the trailing 14-bar average, so quiet chop produces few marks. One pattern max per bar (stars > soldiers/crows > two-bar > single-bar). technicals.<tf>.candle_patterns (last 6) exposes them to the analysis workflow. Caveat: the newest bar may still be forming intraday — a pattern on it can repaint until the bar closes.

123 reversal structures (Sperandeo 1-2-3) are also auto-detected per timeframe from confirmed swing pivots: ① a ~20-bar extreme → ② the reaction pivot → ③ a higher low (bullish) / lower high (bearish). The price pane gets ①②③ markers, a ①→②→③ connector, and a dashed trigger line at the ② price running from ③ to the latest bar; a close beyond ② flips the structure to confirmed (123✓ marker on the breakout bar), while a break of ① drops the structure silently. The ③ marker shows ③? until confirmation. technicals.<tf>.pattern_123 (last 2, each with status: forming|confirmed, trigger, invalidation, p1/p2/p3) exposes them to the analysis workflow — a forming structure is a ready-made entry setup (enter on the ② break, stop beyond ①). The sidebar's 自动信号 section lists them with an 酝酿中/已确认 badge.

Off-session bars (盘前/盘后浅蓝、夜盘深蓝) get a full-height backdrop on both panes — thin-volume price action outside regular hours is visually discounted at a glance. Regular hours = 09:30-16:00 ET (DST-aware via America/New_York).

context — AI-classified news + conclusion (optional, schema_version 2)

Both POST /api/charts (type intraday) and PATCH /api/charts/:id accept an optional context field alongside prediction. It's frozen at write time like prediction — the server never generates or judges it. schema_version is now 2; older (v1) chart docs without context still load and render fine.

"context": {
  "generated_at": "2026-07-06T14:30:00Z",         // ISO timestamp
  "conclusion": {
    "stance": "short",                            // long | short | neutral
    "summary": "一句话综合判断",
    "action": "现在该做什么（挂单/等待/减仓）"
  },
  "news": [
    { "time": "2026-07-06T13:10:00Z",
      "source": "longbridge",                     // longbridge | x | trump | sec | gdelt
      "tag": "catalyst",                          // catalyst | regulatory | sentiment | macro
      "title": "...", "note": "AI 一句话解读", "url": "可选" }
  
    

The dashboard shows context.generated_at's age and a stale badge, sharing the same ~15-min staleness rule as prediction (prediction_updated_at / prediction_stale on chart metas cover both).

Realtime prediction upkeep

Once the US cash session is open, intraday charts must be maintained under these rules:

• Cash-session rebuilds drop off-session bars. Any PATCH after 09:30 ET passes {"session": "intraday", "refresh": true} — the prediction dashboard must not render pre-market / overnight bars intraday. The default --session all is for pre-market analysis only.
• Volume calls align to prior sessions' same-time window. Never compare today's running volume against full-day totals. Pull longbridge kline --period 5m (regular-session bars), sum today's bars, and compare against the same number of opening bars averaged over the prior ~5-8 sessions. Pair the ratio with per-bar direction before calling a move confirmed — a level break on ~0.6x same-period volume is not a confirmed breakout. Caveat: Longbridge daily-K volume includes extended hours; 5m-K volume does not — never mix the two.
• Stale predictions get refreshed on a ~15 min loop. The server marks an intraday chart's prediction stale when it is >15 min old during regular hours (GET /api/charts?stale=true lists them; the SSE envelope and chart metas carry prediction_updated_at / prediction_stale). Each loop round: fetch the stale list → re-pull quote / capital flow / klines → PATCH prediction with scenarios revised only on material change, but always move anchor to the newest m5 bar time + latest price — the anchor marker must track the live tape, never sit minutes behind it → append a timestamped journal note on material revisions (revision discipline) → stop the loop after 16:00 ET close.

桌面版模式

打包的桌面版（Kansoku.app）不监听任何本机端口，全部走 app:// 内部协议 + IPC，不提供 HTTP 接口。要用这个 skill 的 curl 命令，单独起 server 进程即可：pnpm start（或开发模式 pnpm dev），API 照常在 http://localhost:1792。

Storage

• Chart docs: journal/charts/data/<YYYY-MM-DD>-<slug>.json — gitignored, schema_version field for forward compatibility. Date = US session date (derived from the data, not local clock).
• Old single-file HTML archives stay in journal/charts/*.html, listed in the app under 旧版存档 and served at /legacy/<file>.
• The app itself: the pnpm workspace at the repo root (apps/server/ Hono + TS, apps/web/ Vite + React). Analysis parity with the retired Python implementation is locked by vitest golden tests: pnpm test.

Sparkline alternative (no API)

For tiny in-chat previews render Unicode sparklines directly: ▁▂▄▆█ plus ANSI green/red. Use for 5-20-point series where a full chart is overkill.

Related skills

• longbridge-capital-flow / longbridge-kline — same data the server pulls; call directly only for in-chat analysis
• capital-rotation — should end with a cohort chart
• market-session-tracker — may create flow charts
• sepa-strategy — calls sepa as the last step of its Step 10
• intraday-signal — calls intraday twice (POST preview, then PATCH prediction)