dsh-a-share-screener

An A-share stock-screening plugin for DeepSeek Harness (dsh): strategies composed from reusable atomic filters on free, token-less data sources, behind an extensible data-source abstraction.

This is a technical screening tool for historical price/volume patterns. It is NOT investment advice.

Install

Requires the dsh CLI (DeepSeek Harness).

dsh plugin --profile myprofile add github:Gaines-cz/dsh-a-share-screener

Git installs pull source, but the prebuilt lib/ output is committed to the repo, so no build script runs at install time. Pin a commit (github:Gaines-cz/dsh-a-share-screener#<sha>) if you want immutability.

Start with the profile:

dsh --profile myprofile

Use

Ask the agent in natural language:

List the available screening strategies.
Screen all A-shares with low_flat_limit_up, minDrawdownFromHigh 0.7.

Three tools are registered:

Tool	Purpose
a_share_list_strategies	Strategy ids, descriptions, parameters, defaults, valid ranges
a_share_list_filters	Atomic filter ids, descriptions, parameters, defaults, valid ranges
a_share_screen	Full-market scan; returns candidates with quantified evidence

The first full scan downloads history into a local disk cache and can take many minutes (bounded by the data source's rate limit); later scans reuse the cache and only fetch new trade dates. Cancellation is cooperative — aborting the tool call stops the scan.

Built-in strategies

flat_base_low

"Bottom + flat base": the stock trades at or below the 15th percentile of its recent price distribution (~3 years, clamped to available bars) while the last month is a flat, MA-converged base. No limit-up pattern is required. Pair it with minListDays / maxListDays to constrain listing age — e.g. minListDays: 365 + maxListDays: 1460 screens stocks listed 1–4 years.

low_flat_breakout

"Deep low + near-bottom volume breakout" — the right-side twin of low_flat_limit_up: deep below the window high (default ≥ 65% drawdown) while staying near the window low (low ≥ 40 bars back, price ≤ 50% above it), a volume-heavy day (≥ 2× the prior 5-day average) closed ≥ 2% above the high of the preceding one-month flat base with the price holding out of the base since (default ≥ 2 confirming closes), and the MA20 stopped falling with price at or above it. Deliberately no low_percentile gate: a breakout close above the base high sits at the top of the bottom-side distribution by construction.

low_flat_limit_up

"Historical low, flat base, faded volume-heavy limit-up": the stock sits deep below its window high (default ≥ 65% drawdown) at the bottom of its recent distribution (≤ 15th percentile of ~3 years), the last month is a flat, MA-converged base, and within ~6 months there was a limit-up day on ≥ 2× the prior 5-day average volume that has since pulled back below its close while volume cooled to ≤ 40% of the limit-up day.

All thresholds are per-call parameters with defaults (see a_share_list_strategies). Board-aware limit-up thresholds: 10% main board, 20% ChiNext/STAR, 30% BSE. All price-level math runs on a chained daily-return index, so splits and dividends cannot fake a crash or a bottom. Universe filters (all configurable): ST/delisting names, BSE, listings younger than minListDays (default 365) or older than maxListDays (default 0 = no upper bound).

All strategies are compositions of the atomic filters below.

Atomic filters & composition

Every strategy is a declarative predicate over reusable atomic filters, combined with AND / OR / NOT into an expression tree. Fifteen filters ship today (a_share_list_filters lists them with their parameters):

Filter	Gate	Needs
deep_drawdown	latest price ≥ X% below the window high	—
low_percentile	latest price ranks ≤ Xth percentile of the window	—
bars_since_low	the window low lies ≥ N bars back while price stays near it	—
flat_base	recent window is flat with converged MAs	—
platform_breakout	a volume-heavy close cleared the preceding flat base's high and held	—
ma_stabilization	the MA stopped falling (slope ≥ X) and price sits at/above it	—
volatility_regime	annualized realized volatility inside [min, max]	—
volume_limit_up	a volume-heavy limit-up day exists in the window	—
cooldown_pullback	price pulled back below that close and volume cooled off	—
volume_dry_up	absolute volume drought: recent average ≤ X% of the preceding baseline	—
industry_clearance	the industry board itself is in deep clearance (median member drawdown / deep share)	industry (all shipped sources)
industry_position	the industry board sits low in its own range (median member window percentile)	industry (all shipped sources)
market_cap_band	total market cap inside [min, max] 亿元	marketCap (all shipped sources)
amount_liquidity	median daily traded value ≥ X 亿元	amount (eastmoney only)
turnover_band	median daily turnover inside [min, max] %	marketCap (all shipped sources)

Filters that need a capability the active source lacks refuse to run — the scan aborts loudly with the missing capability named instead of silently degrading.

Ad-hoc predicates

a_share_screen also accepts a predicate (mutually exclusive with strategy): a small JSON DSL over the atomic filter ids, e.g.

{ "all": ["deep_drawdown", { "any": ["platform_breakout", "volume_limit_up"] }] }

Groups are { "all": [...] } (AND), { "any": [...] } (OR), { "not": … } (NOT); leaves are filter ids from a_share_list_filters; max depth 3, max 12 leaves. params tunes the composition exactly like a registered strategy. The CLI equivalent is --predicate '<json>'. Strategies using industry_clearance trigger a market-wide aggregation pre-pass; its per-board statistics (median drawdown, deep share, members) appear in each candidate's evidence.

Composition runs one shared derivation pass per stock (chained return index + pre-computed limit-up days), then evaluates the tree: a short-circuit pass for strict hits and a full pass that produces per-gate metrics for the tiered report — which is how "near-miss" candidates (one gate short) are surfaced.

Data source

Three free, token-less sources ship today; sina is the default (its 前复权 closes match market prices).

Source	Token	Notes
Sina (default)	none	前复权 daily bars, 1023 bars per request; the recommended primary
Eastmoney	none	back-adjusted klines per stock (the only source with per-bar amount); clist host fails over realtime → delayed
Tencent	none	后复权 fallback (report prices run high); backup only

The stock list (with industry + market caps) comes from the shared Eastmoney clist endpoint for all three sources.

Every adapter sits behind a DataSource interface (src/datasources/types.ts); the screener, tools, and plugin entry import only that interface, never a concrete vendor. Cache lives under $DSH_HOME/a-share-screener/<source-id>/ (override with cacheDir).

Add a data source

To support another vendor later, implement DataSource and register it in src/datasources/index.ts:

// src/datasources/my-vendor.ts
import type { DataSource } from './types.js'

export function createMyVendorDataSource(limiter: RateLimiter): DataSource {
  async function listStocks(signal: AbortSignal) { /* → StockMeta[] */ }
  async function dailyBars(fullCode: string, startDate: string, signal: AbortSignal) { /* → Bar[] */ }
  return { id: 'my-vendor', capabilities: { industry: false }, listStocks, dailyBars }
}

add myvendor: createMyVendorDataSource to FACTORIES. Set capabilities.industry: true (and populate StockMeta.industry) if the vendor can classify sectors.

Plugin configuration

Set in your profile's cordis.patch.yml (all fields have defaults):

- replace:
    - id: a-share-screener
      config:
        # cacheDir: /path/to/cache   # optional, defaults to $DSH_HOME/a-share-screener
        dataSource: sina             # sina (default) | eastmoney | tencent
        requestsPerMinute: 200
        historyBars: 800
        scanTimeoutMs: 1800000
        excludeST: true
        excludeBSE: true
        minListDays: 365
        maxListDays: 0                # 0 = no upper bound; e.g. 1460 ≈ 4 years

Add a strategy

The preferred way is composing existing atomic filters into a predicate tree — no screening code to write, per-gate diagnosis comes for free:

// src/strategies/my-strategy.ts
import { composeStrategy } from '../engine/compose.js'
import { createFilterRegistry } from '../filters/index.js'

export const myStrategy = composeStrategy({
  id: 'my_strategy',
  description: 'What it looks for, model-facing.',
  // deep drawdown AND flat base (no limit-up requirement this time)
  predicate: {
    kind: 'and',
    children: [
      { kind: 'filter', filter: 'deep_drawdown' },
      { kind: 'filter', filter: 'flat_base' },
    ],
  },
  filters: createFilterRegistry(),
})

For shapes the atomic filters do not cover, implement the Strategy interface by hand (screen returns a hit or null; an optional diagnose powers the tiered report). New atomic filters plug in through src/filters/index.ts with the same Filter interface the built-ins use.

Either way, register it in src/index.ts next to the built-in one — no other changes. The tool schemas and a_share_list_strategies pick it up automatically.

Standalone CLI (no dsh required)

The same code ships a standalone command-line tool: manual trigger, local bar cache, free multi-source data (Sina primary / Eastmoney fallback / Tencent backup), no token.

pnpm install
pnpm sync        # incremental local cache sync (weekly); default full market, narrow with --board / --codes
pnpm scan        # tiered report: strict hits + near-miss candidates, each with gate-level metrics
pnpm strategies  # list strategy ids and their parameter tables
pnpm filters     # list atomic filter ids and their parameter tables
pnpm sources     # list data source ids

Common options: --source sina|eastmoney|tencent, --board <name> (e.g. --board 核能核电), --codes 600519,000858, --strategy <id>, --params k=v,k2=v2, --min-list-days <n>, --max-list-days <n> (0 = no upper bound), --cache-dir <dir>, --out <dir>.

pnpm scan writes reports/<date>-<strategy>-<scope>.md (+ .json). The near-miss tier (exactly one gate failed) exists because a strict multi-gate strategy frequently returns zero hits — it keeps every run reviewable.

Limitations

• The free endpoints (Sina / Eastmoney / Tencent) are public but undocumented; field drift fails loudly rather than silently, and Eastmoney's clist host fails over realtime → delayed so one blocked host does not kill a scan.
• Industry classification is Eastmoney's own taxonomy (f100 on the shared stock list — all shipped sources expose it via capabilities.industry), not the SW (申万) L1/L2 scheme; industry_clearance aggregates on it. Separately, --board screening uses the Eastmoney industry/concept board-member endpoint.
• ST filtering uses the current stock name (no historical name-change tracking).
• Everything runs in the local process; no data leaves your machine except API calls to the data source.

License

MIT