LongPi for DeepSeek Harness

Version 1.1.0. Production bundle, not an MVP slice.

The longevity concierge for DeepSeek Harness: a North-Star biological-age dashboard, five phenotype modules, a medically guarded chat, and a launch-day itinerary — as one installable dsh.bundle.

中文 · Intended use · Security

There is no other DSH plugin that owns this surface. Clinical note-drafting lives in dsh-medseek. LongPi is the member-facing longevity OS: see the number, ask why, take a next step, hand off to a human. Drafts never prescribe.

It is not a medical device. Bundled labs belong to the synthetic member 张明远 (demo).

The claim we can actually defend

Health plugins fail the same way: a number appears on a dashboard and nobody can say where it came from. LongPi's whole pitch is that every number is either recomputable from the repo or explicitly labelled as a demo input.

The headline age is computed, not authored:

npm test   # prints: smoke ok { version: '1.1.0', ... }

test/smoke.mjs recomputes the published PhenoAge model (Levine 2018) from the nine-marker demo panel and asserts the engine lands on 52.52920526313869 years. Change a coefficient, a unit or the Gompertz back-projection and the test fails. That value, plus the five module weights, produces the composite 45.1 against a chronological 45.

Part of the headline	Source	Auditable?
Biological module age	PhenoAge engine, nine blood markers	✅ formula + constants in src/bioage.ts, pinned by a golden test
The other four module ages	Fixed demo inputs	⚠️ labelled demo in every payload, and shown as such in the GUI
Composite	Stated weights, contributions returned	✅ compositeAge() returns each module's contribution
Homeostatic dysregulation	Mahalanobis distance to a young-healthy reference	⚠️ not run without a real reference cohort — the engine refuses instead of inventing one
KDM, Horvath, GrimAge, DunedinPACE	Not implemented	⚠️ listed in read_bioage_model with the reason, not silently omitted

The engine also refuses to score rather than guess: a missing marker, a mg/dL value passed into a mmol/L slot, or a non-physiological age each return a typed failure with the expected unit.

Why this can be the #1 health plugin

The awesome-dsh-plugin catalog is dense in UI chrome, memory, and billing. On longevity / biological age / phenotype, it is empty. The one serious health bundle is MedSeek (clinician SOAP/SBAR + FDA/PubMed). LongPi does not copy that. It ships the missing product:

	LongPi	MedSeek
Who	Member, concierge, launch cohort	Physician / APP / nurse
North Star	Composite biological age + 5 modules	Structured clinical notes
Network	Optional Tycho twin + s2f routing	Allowlisted FDA / PubMed / trials
Guard	No diagnosis, no dose change, 120	PHI egress + citation discipline
GUI	Dashboard tab, radar, provenance line	Theme + settings + search cards

Stars follow a category-defining README, a one-line install, a screenshot whose numbers are true, and the dsh-plugin topic. This repo is built to that bar.

Install

dsh plugin --profile web add github:zwbao/dsh-plugin-longpi
dsh web

From a checkout:

npm install && npm test
dsh plugin --profile web add "$PWD"
dsh --profile web --dump-config   # expect "# == dsh-plugin-longpi"

If dsh is not on PATH:

npx -y --package @deepseek-ai/dsh dsh plugin --profile web add "$PWD"

Restart the GUI (or hard-refresh). Open a session → tab LongPi. Uninstall: dsh plugin --profile web remove dsh-plugin-longpi.

What you get

Surface	What you see
Conversation view LongPi	Composite age, Δ vs chronological, 5-module radar, status chips, insights, Oct 24 itinerary
Composer dock	Suggested questions (copy → paste into Chat)
Sidebar mark	LongPi dot
Persona	Health assistant, not a coding agent
Tools (23)	8 concierge + 4 biological-age + 11 genome/omics/s2f — matches source
Skills (6)	Phenotype, panel, itinerary, genome, multi-omics, s2f routing

Tools (23)

Tool	Answers	Writes
read_dashboard	Composite age, five modules, insights	no
read_phenotype	One module + metrics	no
explain_metric	One demo metric vs reference / previous	no
list_insights	Reviewed insight cards	no
get_event_briefing	2026-10-24 itinerary	no
search_faq	Keyword FAQ + glossary	no
create_appointment_request	Concierge intent (max 1 / process)	yes
handoff_concierge	Human handoff (max 1 / process)	yes
compute_biological_age	PhenoAge from nine markers; refuses on missing or mis-united input	no
read_bioage_model	Model card: citations, required units, claim ceilings, what is not implemented	no
read_demo_lab_panel	The synthetic nine-marker panel and the ages it produces	no
compute_composite_age	Weighted composite + each module's contribution	no
s2f_route	Rank s2f-agent skills	no
s2f_plan	Dry-run plan + missing canonical inputs	no
s2f_execute	Optional penguin s2f route (off by default)	no
s2f_batch_request	De-identified JSON for s2f batch (gpn_msa / translate)	no
read_personal_genome	12-locus longevity panel	no
ingest_vcf	Local VCF SNP ingest (hg38, capped)	in-process
annotate_variant	Consequence, ClinVar note, literature, s2f skills	no
annotate_multiomics	Five omics layers	no
build_omics_report	1.1.0 combined report	no
export_report	JSON or Markdown	no
lookup_longevity_evidence	Curated index with verified licence and integration mode per row	no

Skills (6)

• phenotype-literacy — composite age and module weights
• panel-interpretation — hs-CRP, inflammatory age, VO₂ max, HRV, sleep
• event-day-itinerary — launch-day path
• genome-personalization — demo WGS panel + s2f scoring path
• multiomics-annotation — layered omics, never fill missing layers
• s2f-routing — s2f-penguin guards + profile batch contract

Quick start

1. Install, restart dsh web, new session, open the LongPi tab.
2. Ask in Chat: “我的 hs-CRP 高吗？” The agent should call explain_metric / search_faq and quote 4.2 mg/L (demo).
3. Ask: “10 月 24 日怎么走？” → get_event_briefing.
4. Try a medication-change phrasing. The guard must refuse and point to a concierge.

Every answer that cites a number must come from a tool or the demo customer_block. Invented labs are a bug.

Architecture

flowchart LR
  subgraph dsh["DeepSeek Harness"]
    WEB[Web GUI]
    TOOLS[Tool registry]
    SK[Skill registry]
    SP[System prompt]
  end
  subgraph bundle["dsh-plugin-longpi"]
    HOST[Host: tools + guard + persona]
    FIX[Synthetic fixture 张明远]
    UI[Client: Dashboard + dock]
    API["GET /api/longpi/dashboard"]
  end
  WEB --> UI
  UI --> API --> FIX
  HOST --> TOOLS
  HOST --> SK
  HOST --> SP
  HOST --> FIX

One package, one cordis.patch.yml row. No EHR. No wearable OAuth. No third-party Cordis plugins inside the customer process.

Where LongPi stops, and what it delegates

LongPi is deliberately the thin part of a three-part system. It owns the member-facing surface and the frontier evidence; it does not re-implement the databases or the longitudinal statistics that already exist and are better than anything this repo would write.

Layer	Owner	What it holds	How LongPi talks to it
Member surface + frontier evidence	LongPi (this repo)	Biological-age computation, the nine-marker panel, genomics/omics annotation, s2f routing, guardrails, GUI	—
Sequence-to-function claims	s2f-penguin / s2f-agent	Variant-effect model adapters, claim ceilings, run receipts	s2f_route / s2f_plan / s2f_batch_request emit a de-identified contract; execution stays in that repo. DSH never runs GPU inference.
This person's long-term data	Tycho Engine (local, MCP)	Hash-chained ledger, twin labs/genotype, forecast resolution and calibration, N-of-1 sequential analysis, safety gate, screening loop	One @deepseek-ai/dsh-mcp-client row in cordis.patch.yml (shipped commented out). Tools appear as mcp__tycho__*.

The boundary rule: LongPi proposes candidates with an explicit claim ceiling; Tycho decides what is true about this individual. A phenotype-age number never becomes a treatment instruction, and a forecast never becomes a conclusion before its window resolves.

To turn on the twin, follow the annotated block at the bottom of cordis.patch.yml (tycho paths → tycho list-tools → tycho serve).

Safety model

Layer	Mechanism
Banner	Persistent 演示数据 / DEMO DATA
Persona	No 诊断 / 患有 / 治愈; no dose changes
agent/pre-step	Emergency → 120 script; medication intent → refuse
Write caps	1 appointment, 1 handoff per process
Network	Dashboard API is local; tools do not fetch
Refusal over guessing	Missing marker → MISSING_INPUT; implausible unit → OUT_OF_RANGE; no reference cohort → NO_REFERENCE
Provenance	Every engine result carries model, citation, units, claim ceiling; the GUI marks computed vs demo inputs
Licence honesty	lookup_longevity_evidence labels each project import / process / cite / unknown; copyleft rows can never claim import (asserted in the smoke test)

Details: docs/intended-use.md, SECURITY.md.

Configuration

cordis.patch.yml:

config:
  brandName: 健康助手
  demoBanner: true
  itineraryDate: '2026-10-24'

Compatibility

dsh	tested on @deepseek-ai/dsh 0.1.5-rc.1
Node	^22.19 || >=24

Development

npm install
npm test          # tsdown + smoke (PhenoAge golden vector, tools, guards, VCF, report)

Use Node 22 or 24. The bundler (rolldown) ships a native arm64 binding; on Node 20 npm test fails with Cannot find module '@rolldown/binding-darwin-arm64'. If your default node is older:

PATH="$(brew --prefix node@22)/bin:$PATH" npm test

npm test asserts the PhenoAge golden vector 52.52920526313869 (recomputed from the nine-marker panel), that a missing marker yields MISSING_INPUT, that 100 in the glucose slot yields OUT_OF_RANGE, that HD without a reference cohort yields NO_REFERENCE, that all 23 tools register, hs-CRP alias crp → hs_crp, medication and emergency guards, and FAQ retrieve.

Listing checklist (awesome-dsh-plugin)

• dsh.bundle + cordis.patch.yml
• dsh plugin add installable
• MIT LICENSE
• Tool count in README matches source
• GitHub topic dsh-plugin
• One-line description on the list, category Tools & Capabilities or Skills

License

MIT. The dsh-plugin topic is a discovery tag, not a DeepSeek endorsement and not a medical clearance.