📊 dsh-fund-research

Deterministic research reports for Chinese public mutual funds, on DeepSeek Harness.

Every key number in every report traces back to a hashed source snapshot — gaps declared, never invented. Research only; not investment advice.

English · 简体中文 · Español · Português · हिन्दी

Compatibility

What you get

• fund_research tool — one fund code in, a versioned Markdown research report out: overview, performance decomposition, holdings penetration, simplified style attribution, manager profile, risk & gap declarations, disclaimer, and a number-traceability appendix mapping every key figure to its snapshot JSON path and verification verdict. Sealed to fund-reports/{code}/{YYYYMMDD-HHmmss}/ as report.md + manifest.json + snapshot.json. background: true runs it as a fund-report background job.
• fund_snapshot tool — a light snapshot card (latest NAV, published stage returns, scale, manager, top-3 holdings) sealed into the fund's day directory.
• Deterministic metrics, zero model arithmetic — period/annualized return, volatility, max drawdown, Sharpe; top-N concentration, HHI, industry distribution, quarter-over-quarter holdings comparison; size-value style bands; manager tenure and peer comparison. All pure functions over the sealed snapshot.
• Traceability as a first-class feature — before sealing, every key number is checked against the sealed snapshot.json through the optional dsh-data-quality service when it is installed, or through the built-in isomorphic fallback checker (builtin-fallback) otherwise. The appendix table records value ↔ path ↔ verdict.
• Honest gaps — a failed or degraded data source produces an explicit 数据缺口 (data gap) declaration in the affected section. The plugin never fills a gap with an invented number.
• Offline mode — offline: true (config or tool argument) serves everything from the storage-domain snapshot layer or the newest on-disk version snapshot, with zero outbound requests. Ideal for tests and reproduction.
• asOf cutoff — asOfDate (ISO YYYY-MM-DD) truncates the NAV series to data on or before that date and stamps the snapshot + report with the cutoff; invalid or future dates fail loudly.
• Checkpoint resume — <reportRoot>/.run-state.json records each pipeline stage (snapshot/report) with timestamps and an input fingerprint; resume: true continues from the first incomplete stage, reusing sealed artifacts, and rejects a fingerprint mismatch.
• Source discovery record — every acquisition seals a code-generated sources-discovery.json (endpoint roster, primary/fallback resolution, per-source coverage and gaps, degradation reasons) and folds it into the report appendix as 数据源与缺口声明.
• Multi-fund fan-out — codes accepts an array of fund codes; each fund runs the pipeline independently with per-fund failure isolation (failures become summary gaps), and the result is a summary card (code / asOf / seal hash / verdicts / failure reason).
• Tracking ledger — every successful seal appends a deterministic line to <reportRoot>/.tracking.jsonl; includeComparison: true renders a deterministic 与上次对比 section (NAV range / scale / top holdings) with a gap declaration when no prior record exists.
• Read-only review — after sealing, a fund-review job reviews the sealed artifacts (gap-declaration completeness, traceability-table consistency, disclaimer) and writes review-note.md; it skips gracefully (recorded in run-state) when no jobs service is present.
• Per-source quality signals — every source carries deterministic quality metadata (requested/succeeded/fieldsPresent/parseWarnings/degraded), rendered in the appendix and surfaced in tool values so downstream can downweight (never hard-filter) a low-quality source.
• Walk-forward stability summary — includeWalkForward: true adds a 样本外稳定性摘要 section: deterministic rolling-window return/Sharpe sign persistence and mean/std, explicitly labelled as statistical description only, not a prediction.
• Session audit events — fund-research/snapshot and fund-research/report log-only events carry the code, version directory, manifest hash, and gap list (model-visible ⟺ logged).
• Methodology skill — a bundled fund-research skill teaches the model the metric口径 (definitions), gap handling, and compliance wording. Computation stays in code.

Quick start

> 用 fund_research 出一份 161725 的研究报告

The agent calls fund_research({ code: "161725" }); a minute later the workspace holds:

fund-reports/161725/20260819-153012/
├── snapshot.json            # raw extracted data + computed metrics + per-source sha256
├── sources-discovery.json   # code-generated endpoint roster + coverage + gaps
├── report.md                # the research report with the traceability appendix
└── manifest.json            # snapshot/report hashes, parameters, verify engine, gaps

.run-state.json sits at the report root and records the pipeline stages for resume: true. Every number in report.md's appendix carries a verified / mismatch / not-found / unverifiable verdict against snapshot.json — recompute any of them from raw.* with the documented口径 to audit the plugin itself.

Install & uninstall

dsh plugin --profile web add dsh-fund-research     # install (npm or tarball)
dsh plugin --profile web remove dsh-fund-research  # uninstall

Restart the profile after installing (bundle activation is restart-based). The bundle patch composes the storage stack (dsh-storage + dsh-storage-json + dsh-storage-domain) the snapshot layer needs.

Configuration

All keys are optional (defaults shown); invalid values fail loudly at load.

Tools & surfaces

fund_research

fund_snapshot

Report sections

概览 overview · 业绩拆解 performance decomposition · 持仓穿透 holdings penetration · 风格归因 style attribution (simplified) · 经理画像 manager profile · 同类/指数基准对比 benchmark & peer comparison · 风险与缺口声明 risk & gaps · 免责声明 disclaimer · 附录：数字回溯表 traceability appendix.

Permissions & data

• Reads the public Tiantian Fund / Eastmoney endpoints (fund.eastmoney.com/pingzhongdata/*.js, fundf10.eastmoney.com F10 pages, push2.eastmoney.com quotes) with a browser User-Agent and configurable polite pacing. No key, no login, no paid API, no anti-crawler circumvention.
• Writes only under the configured report root inside the session workspace, plus the dsh_fund_research storage domain (latest snapshot per fund).
• Never evaluates remote JavaScript (the pingzhongdata block is scanned, never executed), never stores credentials, never trades.
• Session events are log-only audit records; the pinned 0.1.1-rc.2 peers offer no ignorable envelope, so a session restored by a build without this plugin refuses those log lines — the same accepted trade-off as other research plugins of this family.

Security boundaries

• Fund codes are validated as exactly six digits before touching a path or a URL; the report root resolves inside the session workspace.
• Source payloads are hashed (SHA-256) at acquisition; the sealed manifest lets you detect silent upstream edits between runs.
• Verification never blocks a seal: a broken optional dsh-data-quality service degrades to the built-in checker, and the engine used is recorded in the manifest and the appendix.
• See SECURITY.md for the reporting policy.

Known limitations

• Upstream structure drift. The parsers are strict by design: if Tiantian Fund changes a var Data_* shape or an F10 table layout, the affected source throws a SourceParseError naming the field, and the section degrades to a declared gap (the core pingzhongdata block failing aborts the run loudly). This is deliberate — a silent misparse is worse than a declared gap.
• Style attribution is估算口径. Fixed size bands (≥1000亿 / 300–1000亿 / <300亿) and PE bands, plus within-holdings quintiles — no full-market distribution is consulted. The report labels this.
• Holdings are quarterly disclosure data (披露滞后); the F10 page carries the latest two quarters.
• One fund per call; no portfolio analysis, no PDF annual reports, no real-time quotes (the fundgz.1234567.com.cn realtime endpoint is dead and deliberately unused).
• The Web UI "deliverables" turn row keys off mutation-tool call cards; this plugin's produced files surface through the tool call card's follow-along location (the fund's report directory), not per-file rows.

Development

pnpm install
pnpm run typecheck && pnpm run typecheck:ci   # types, incl. CI-strict
pnpm test                                     # 124 tests over real harness seams
pnpm run test:e2e                              # opt-in LIVE-network E2E (LIVE_E2E=1)
pnpm run build && pnpm run verify:artifacts   # tsdown + tsc declarations
pnpm run verify:self-contained                # no out-of-repo dependency specs
node scripts/check-readme-sync.mjs            # five-language README gate
node scripts/check-endpoints.mjs              # M3 endpoint-liveness probe (4 eastmoney hosts)
pnpm pack                                     # tarball

Tests run the REAL Context/SessionStore/ToolRuntime/LocalJobRegistry/storage seam from the 0.1.1-rc.2 peers; the network is replaced only at the fetch boundary by saved real-response fixtures (fixtures/, fund 161725). Refresh fixtures with the collector scripts in .tmp/.

Topics

dsh · dsh-plugin · deepseek-harness · cordis · fund-research · mutual-fund · investment-research · finance · research-report

Contributors

• PerryLink — maintainer: the collector/metrics/report-seal pipeline, the endpoint-liveness probe, CI and releases, and the five-language docs.
• dsh-fund-research contributors — collective author of the foundational build (plugin contract, config schema, tools, tests, packaging).

No external contributors yet — 0 community PRs/issues merged. Open an issue via the forms in .github/ISSUE_TEMPLATE/ or a pull request against main to be listed here.

PerryLink DSH Plugin Family

Part of a family of standalone DeepSeek Harness plugins sharing one engineering baseline: pinned 0.1.1-rc.2 peers, fail-loud Schemastery config, five-language READMEs, and real-seam vitest coverage.

License

Apache-2.0. Third-party notices: THIRD_PARTY_NOTICES.md.

Disclaimer: this plugin produces research artifacts only. Nothing it outputs constitutes investment advice.