CLI & agent skill for a PayPay証券 (PayPay Securities, ペイペイ証券) account: check portfolio, holdings, balance, 投資信託, 米国株, 取引履歴 (transaction history), fees & FX-spread; generate a 復盘/review with realized & unrealized P&L and money-weighted return (XIRR); 持仓结构/exposure (risk), 定投/つみたて recurring-buy plans (plans), per-year tax view (tax), account snapshots & diff (snapshot/diff), multi-account consolidation (-a all), Chinese output (--lang zh); AND place 米国株 orders (buy/sell/cancel). Use when the user wants to view, review, analyze, report, or trade a PayPay証券 / PayPay investment account. Orders dry-run by default; a live order requires an explicit human `--execute` with a typed confirmation + the account TRADE_PASSWORD (the agent never auto-submits). Facts only — never investment advice or buy/sell judgments.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
paypay-securities
description
CLI & agent skill for a PayPay証券 (PayPay Securities, ペイペイ証券) account: check portfolio, holdings, balance, 投資信託, 米国株, 取引履歴 (transaction history), fees & FX-spread; generate a 復盘/review with realized & unrealized P&L and money-weighted return (XIRR); 持仓结构/exposure (risk), 定投/つみたて recurring-buy plans (plans), per-year tax view (tax), account snapshots & diff (snapshot/diff), multi-account consolidation (-a all), Chinese output (--lang zh); AND place 米国株 orders (buy/sell/cancel). Use when the user wants to view, review, analyze, report, or trade a PayPay証券 / PayPay investment account. Orders dry-run by default; a live order requires an explicit human `--execute` with a typed confirmation + the account TRADE_PASSWORD (the agent never auto-submits). Facts only — never investment advice or buy/sell judgments.
paypay-securities — read-only PayPay証券 client
A small Python CLI that authenticates to the PayPay証券 web frontend and reads
account data. There is no official or unofficial API/SDK for PayPay証券, so
this talks to the same endpoints the website uses.
Scope
Phase
Status
What
0 — auth
✅ done
/login.json + trusted-device cookie → session token, no SMS
1 — read (証券)
✅ done
balance, portfolio (per-holding), asset/cash history
1b — read (投信)
✅ done
mutual-fund valuation/principal/P&L via JSON API; total aggregation
2 — export/notify
not built
scheduled pull + push to Lark/WeChat/email
3 — trading (証券 米国株)
✅ built
buy/sell/cancel via the web order flow — dry-run by default; live --execute is human-only (typed confirm + TRADE_PASSWORD); per-order + daily-count guards
Account is split across systems (see also the project memory): 証券 (this
site, incl. the cash balance in the transaction ledger), 投信 (/investment_trust/
Vue SPA + JSON API, same session), and CFD (cfd.paypay-sec.co.jp, separate login
— not integrated). paypay total/assets cover 証券 + 投信 + cash = the app's
full grand total; only CFD is out of scope.
The CLI (the bundled paypay_sec/ package + pyproject.toml) is installed
alongside this SKILL.md. Requires uv on PATH.
Setup (credentials)
Put credentials in ~/.paypay-sec/.env (outside the repo / installed skill —
this path is searched first, so the CLI works from any directory). Template in
.env.example:
PAYPAY_MEMBER_ID, PAYPAY_PASSWORD
PAYPAY_COOKIE — the full Cookie header from a logged-in browser. It must
contain the ..._SMS_AUTH_STRING trusted-device token, otherwise the server
demands an SMS code and login fails.
Never commit credentials or pass them on the command line. (PAYPAY_ENV overrides
the path; ./.env / ./spike/.env are also searched for dev checkouts.)
Multiple accounts: each account is a profile. The default account reads
~/.paypay-sec/.env; a named account <name> reads ~/.paypay-sec/<name>.env.
Select one with -a <name> (or the PAYPAY_ACCOUNT env var) — e.g.
uv run paypay assets -a second. Each account keeps its OWN session + response
cache (default → ~/.paypay-sec/, named → ~/.paypay-sec/<name>/) so they never
collide. paypay accounts lists the configured profiles.
Commands
Run from this skill's directory: uv run paypay <cmd> (uv resolves deps from the
bundled pyproject.toml; the paypay entry point is defined there).
uv run paypay login # force a fresh login (refreshes the cached session)
uv run paypay logout# clear the cached session
uv run paypay balance # 証券: 評価額合計 / 投資元本 / 含み損益
uv run paypay portfolio # 証券 holdings: valuation, shares, cost, P&L, account type
uv run paypay history# 証券 daily asset/cash time series
uv run paypay invtrust # 投信 (mutual funds): valuation / principal / P&L / 売却申込中
uv run paypay invtrust-history # 投信 ledger (MARKET_ID=99): 買付/売却/入金/譲渡益税/送金手数料 + moving-avg realized P&L + 現保有(口座別 NISA/特定)
uv run paypay total # aggregate 証券 + 投信 invested assets (excludes cash)
uv run paypay assets # one-shot consolidated holdings + cash + grand total (parallel)
uv run paypay trades [--pages N] # 証券 transaction ledger (買付/売却/入金/手数料) + running cash balance
uv run paypay fees [--detail] # cost analysis: explicit fees + measured FX spread (+ optional price spread)
uv run paypay review # 复盘: 2 blocks — 持仓盈亏(評価損益=App頭条) + 账户盈亏(通算=総資産−純入金); + realized/costs
uv run paypay trades-summary # per-brand buy/sell/net-invested/net-shares/realized P&L
uv run paypay risk [--accounts] # 持仓结构 / exposure: weights, concentration, FX/category split (+口座区分 with --accounts; FACTS ONLY)
uv run paypay plans # 定投/つみたて: active recurring-buy funds + monthly run-rate (scans FULL history by default; --fast = quick)
uv run paypay tax # per-year tax view: 売却 / 譲渡益税 / 分配金 (FULL history by default; --fast = quick; 年間取引報告書 参考, FACTS ONLY)
uv run paypay snapshot save # save a dated account snapshot (the CLI's own time series)
uv run paypay snapshot list # list saved snapshots
uv run paypay diff [--days N] # diff a live read vs the latest (or ~N-days-old) snapshot
uv run paypay calendar # build 每日涨跌日历 HTML from snapshots (--json = raw schema; --out PATH; -a all = household, the default)
uv run paypay doctor [--online] # diagnose setup/login readiness (offline by default; --online probes login + 証券/投信 to confirm the session is live)
uv run paypay accounts # list configured account profiles
uv run paypay cache-clear # clear the local response cache
risk is facts-only. It reports portfolio structure — total, cash %, largest
position %, top-1/3/5 concentration, 種類/口座 splits, and two distinct currency/region
measures: 米国 計価(証券, USD-quoted) vs 米国株 底層暴露 (US-underlying, which also
counts JPY-priced S&P500/NASDAQ etc. 投信). It gives no risk verdict and no buy/sell
advice, same boundary as every other command.
snapshot / diff give the account its own long-term series:
snapshot save writes <state_dir>/snapshots/<ts>.json (assets, cash, holdings,
realized, deposits); diff compares a live read against the latest snapshot (or
--days N ago) so you get "this week's change" — asset/holdings/deposit/realized deltas.
Automate it: bin/snapshot-cron.sh install schedules a daily (07:30 local, trading
days) read-only snapshot of every profile + a calendar rebuild + a Monday
diff --days 7, logged to ~/.paypay-sec/snapshot-cron.log (uninstall / status too).
calendar reads only the saved snapshots (no creds/network) and builds a
single self-contained, offline HTML — the 每日涨跌日历: month + year (GitHub-style
heatmap) views, 红涨/绿跌, per-account toggle + 合计. Each day's value is the
snapshot-to-snapshot change in 総資産 minus that day's net cash flow (so deposits
aren't counted as profit; full-quality, includes 投信). Needs ≥2 snapshots on
different days per account. Default / -a all = household (all profiles); -a <name>
= one. Labels come from PAYPAY_LABEL in each .env (else the profile name). The
template (paypay_sec/calendar_template.html) is data-driven: the cron regex-swaps
the JSON between /* DATA_START */ … /* DATA_END */. --json emits the raw schema.
Any command takes -a <name> to target a non-default account, and
--format table|lark|json. --format lark emits Feishu/Lark-friendly
bullets (bold numbers, +¥/-¥, no wide tables) — use it for review /
trades-summary / risk / diff when posting to Lark.
--lang ja|zh localizes the table/lark labels (zh = 中文: 総資産→总资产,
評価損益→持仓盈亏, 実現→已实现, …); JSON output keys stay English. --all pages
the full ledger history (until NEXT_FLG=false) for complete realized P&L instead of
the default page cap. Reports also stamp 查询时间 (as_of) + per-source freshness
(live/cache/stale) and loudly flag any feed that failed — never silently ¥0.
JSON is versioned: every --json dict payload carries schema_version (currently
"1.0") so cron jobs / dashboards / the daily snapshot can parse it stably.
-a all consolidates across every configured account profile (total / assets /
risk / plans / tax): combined household view (same fund summed) + per-account
totals — e.g. risk -a all shows whole-household concentration & US-underlying exposure.
review also
reports a money-weighted return (XIRR, 資金加重収益率) — the proper annualized
performance when 定投/deposits are ongoing (use --all for the full deposit history).
Scope of the read commands: data display only. The commands above just show
your account's data (or factual calculations on it — totals, realized P&L by
average-cost basis, cost aggregation). They give NO judgments, NO risk/position
rules, NO buy/sell advice. Order placement is a separate, explicit flow ↓ — and
it too gives no advice: it places exactly the order you specify, nothing more.
Ordering (下单 — 米国株 buy/sell/cancel)
Trading is OFF by default.buy/sell/orders/cancel refuse to run unless
the human sets PAYPAY_TRADING_ENABLED=1 (in ~/.paypay-sec/.env or the shell).
A read-only / 复盘 invocation can't place an order even by accident — this capability
gate sits on top of the dry-run + confirm + TRADE_PASSWORD wall below.
PayPay証券 has no order API; this drives the same un-pinned web order endpoints
the site uses (/trade/brand/ajax_*_popup → ajax_*_complete). The agent never
auto-submits a live order. Every order is a dry-run (見積/preview only) unless a
human adds --execute, and --execute then requires (1) typing an exact
confirmation phrase and (2) entering the account TRADE_PASSWORD (取引パスワード)
at an interactive getpass prompt. The agent does not know, store, or handle that
password — only the live confirm/preview ever runs unattended.
# DRY-RUN (default) — runs the real 見積/confirm, prints the quote, places NOTHING
uv run paypay buy TSLA --amount 10000 # 金額指定: ¥10,000 of TSLA, at the quote
uv run paypay buy TSLA --qty 1 # 株数指定: 1 share
uv run paypay sell TSLA --qty 0.5 # sell 0.5 share
uv run paypay buy QQQ --amount 50000 --account-type 3 # into 成長投資枠 NISA# LIVE — human only. Prompts for the confirmation phrase, THEN the TRADE_PASSWORD.
uv run paypay buy TSLA --amount 10000 --execute
uv run paypay orders # list 未約定/予約注文 (pending orders)
uv run paypay cancel <ORDER_ID> # cancel a pending order
Guards (in guards.py / TradeConfig): allow_markets (米国株 only for now),
max_order_jpy per order, and a daily order-count cap. A blocked order prints
⛔ order blocked by guards: … and never reaches the network confirm.
Every order attempt is appended to an audit log (audit.py) with a daily count.
Funding model: the account is funded by cash balance (you top it up); orders are
placed against that cash. Bank-card buying is mobile-app/passkey-only and out of scope.
Run from anywhere (no cd): symlink the bundled launcher onto your PATH —
ln -s <skill-dir>/bin/paypay ~/.local/bin/paypay (skill-dir varies by runner:
~/.claude/skills, ~/.hermes/skills, ~/.agents/skills, …) — then
paypay review --format lark works in any directory (the launcher resolves the
symlink + skill dir and runs uv run there).
total / assets scope: 証券 (株+ETF) + 投信 holdings + the account cash
balance, giving the full grand total that matches the app's 保有資産 figure. The
cash comes from the transaction ledger's running CASH_BALANCE
(/trade/history/ajax_settlement.json), NOT from PayPayマネー — the cash lives on
the securities site after all. Only CFD (cfd.paypay-sec.co.jp, separate login)
is excluded.
Response cache (anti-throttle): every GET/POST response is cached to
~/.paypay-sec/cache/ with a TTL (default 120s, env PAYPAY_CACHE_TTL). Repeated
or overlapping commands (and the total/assets/trades trio, which share the
settlement ledger) reuse cached data instead of re-hitting the API — important
because the ajax_settlement endpoint throttles (returns empty) when hammered.
--no-cache forces fresh fetches; paypay cache-clear empties it.
Session reuse: after the first login the session cookies + token are cached
to ~/.paypay-sec/session.json (mode 0600) and reused, so /login.json is hit
only on a cold start or when the session has expired (a fetch bouncing to
/login/ triggers exactly one automatic re-login). This keeps load off the
login endpoint instead of authenticating on every command.
Flags (place AFTER the subcommand, e.g. uv run paypay portfolio -m usa -a second --json):
config.py — credential loading from .env/env (never logged).
client.py — PayPayClient: login() + get_page(); market-slug aliases;
the web order methods order_confirm() / order_submit() / open_orders() /
order_cancel() (FuelPHP CSRF from the fuel_csrf_token cookie; ticker→BRAND_ID
resolution; single-use ORDER_CONFIRM_NO tokens).
orders.py — order model (OrderRequest) + the build→guard→confirm→[confirm
phrase]→submit pipeline (dry_run() / place()); injected confirm/submit
callables so the live submit is never reached except via the human path.
guards.py — TradeConfig + check_guards(): allow-listed market, per-order
yen cap, daily order-count cap (pre- and post-confirm checks).
audit.py — append-only order log + daily order count.
parsers.py — all site-specific selectors live here. Data sources:
holdings → table.d_table on /trade/portfolio/brands/<market>
history → var ticks/cashData/acuisitionData JS arrays (values in 万円)
投信 valuation → parse_invtrust() over the POST /v2/invest/brand/pc_invest_top
JSON (NOT HTML — the 投信 page is a Vue SPA; the JSON API needs the FormData
fields APP_VERSION/UUID/DEVICE_TOKEN/OS/APP_ID, empty body hangs the server)
投信 ledger → parse_invtrust_transactions() over
GET /v0/history/settlements.json?MARKET_ID=99&PAGE_NUM=<n>&OS=pc — a SEPARATE
market view from the 証券 ajax ledger (the SPA's "取引の履歴" tab). Holds 投信
買付/売却/入金/譲渡益税(type 8)/送金手数料(type 54), with 口座区分 in
ACCOUNT_TYPE (1 一般 / 2 特定 / 3 NISA成長 / 4 NISAつみたて). PAGE_NUM here is a
RECORD OFFSET (n..n+19), same gotcha as the 証券 ajax — step by 20 + de-dup by
SEQ_NO (paging by 1 without de-dup inflates totals ~10×). MINI_CLIENT_SEQ_NO
is optional. The 入金/送金手数料 rows are account-wide (identical to the 証券
ledger); only 買付/売却/譲渡益税 are 投信-specific. review folds its realized
P&L + 譲渡益税 + 送金手数料 into the full-account 総収益 so the 証券-only residual
no longer hides 投信 trading. The endpoint throttles (empty) when hammered.
open orders → parse_open_orders() over table.d_table on /trade/preorder/
market.py — real USD/JPY mid (ECB via frankfurter.dev), cached forever on disk.
costs.py — paypay fees logic: explicit fees from the ledger + FX spread
measured as applied 為替レート vs market mid. PayPay never itemizes its spread
(ledger and 取引報告書 PDFs both show 手数料=¥0; cost is baked into 約定価格 +
為替レート), so the FX-spread number is reconstructed, and the price spread
(~0.5%/0.7%) is only an optional --price-spread-pct estimate.
report.py — review / trades-summary / invtrust-history aggregation:
deposits, per-brand buy/sell/net, realized P&L via moving-average cost
(移動平均法 — what JP 特定口座 uses; correct when buys & sells interleave, unlike
a whole-window average). BrandFlow.events holds the chronological trade stream;
callers feed it oldest-first (reversed(ledger)). A reconciles flag goes False
when a sell exceeds the basis on record (window too short → realized under-est.).
Pure facts — no rules/judgments.
cli.py — argparse subcommands + rendering (CJK-width-aware tables; table/lark/json).
The order subcommands wire the pipeline, prompt for the confirmation phrase +
TRADE_PASSWORD (getpass, never stored), and emit the dry-run/submitted/blocked
result. _persist_cash() caches 現金 to ~/.paypay-sec/[acct/]last_cash.json and
falls back to it (marked ⚠stale) when the settlement ledger throttles to empty,
so assets/total/review never silently report cash as ¥0.
Dev-only, NOT shipped: repo-root tests/ (real-fixture tests; gitignored as it
holds account HTML/JSON). Run: uv run --project skills/paypay-sec python tests/test_parsers.py.
Committable synthetic regression tests for the 投信 ledger + moving-average:
skills/paypay-securities/selftest.py —
uv run --project skills/paypay-securities python skills/paypay-securities/selftest.py.
Fragility: the 証券 pages are server-side-rendered HTML, so a frontend
redesign breaks parsing — fix only parsers.py (all selectors live there).
Safety
The read commands are read-only. The order commands place trades, behind a
human-in-the-loop wall:
Trading is disabled by default.buy/sell/orders/cancel refuse to run
unless PAYPAY_TRADING_ENABLED=1 is set — a read-only/复盘 session has no order
capability at all.
Dry-run is the default. Even when enabled, a bare buy/sell only runs the
見積/preview; it places nothing. A live order requires an explicit --execute.
--execute is human-only. It demands a typed confirmation phrase and the
account TRADE_PASSWORD at an interactive prompt. The agent never enters that
password and is not built to run a live submit unattended.
Guards + audit. Per-order yen cap, allow-listed market, daily order-count
cap; every attempt is logged. Confirm tokens (ORDER_CONFIRM_NO) are single-use
and writes are never auto-retried.
No advice. The skill places exactly the order you specify and never offers
buy/sell/position judgments.
Automated access to a brokerage may conflict with PayPay証券's terms of service —
use on your own account at your own risk. The order confirm/submit response field
names + the open-orders column mapping are finalized against a live capture when
the US market is open (closed → buyable=0 → confirm returns STATUS=false), so run
a dry-run buy at market open to validate before the first live --execute.
