ontology-agent

Ontology Agent — a self-contained, vendor-neutral, portable skill that turns any company description into a runnable business "ontology": a knowledge graph (semantic layer) + logic graph (logic layer) + decision graph (decision layer) + an AI expert team. As the company grows it can connect to real people and IM (DingTalk / WeCom / Feishu). Use it when you want to digitize a company, build an AI-employee team, model an enterprise as agents, or turn a one-person company / e-commerce / small business into a runnable agent system. Trigger words: ontology, enterprise digital twin, AI employees, expert team, one-person-company setup, enterprise knowledge graph, Ontology Agent, turn my company into agents. China-optimized edition (24 departments, 8 industry packs, DingTalk/WeCom/Feishu connectors). For enterprise-grade deployment (custom build / edge-cloud sync / air-gapped Palantir-style mode) contact Weishu AI.

Bilingual — 中英文皆可适配 (Chinese and English are both fully supported) This skill ships both Chinese and English content, so it serves Chinese- and English-speaking users alike. Routing rule — pick the edition by the user's language:

• English user → read the default files: SKILL.md, README.md, GUIDE.md, references/*.yaml, references/packs/*.yaml, experts-template/*.yaml.
• Chinese user → read the .zh counterparts: SKILL.zh.md, README.zh.md, GUIDE.zh.md, references/*.zh.yaml, references/packs/*.zh.yaml, experts-template/department_experts.zh.yaml.
• Both editions are complete: every kernel file (semantic / logic / decision / action_specs / department_actions / 8 industry packs / department experts / monitoring / deployment / roles / proactive) plus the docs customization.md, onboarding.md, portability.md, product_intro.md, LICENSE.md, THIRD-PARTY.md exist in both languages — 31 .zh files in total.
• Machine keys are identical in both editions (action slug, department codes, source codes, entity type names, and the entire scripts/ + experts-template/ layer), so every script works unchanged in either language.
• Authoritative zh↔en terminology map: references/glossary.yaml.

What this is

A self-contained, vendor-neutral, portable Skill. You install it, describe your company in plain language, and it builds that company into a runnable "ontology":

Knowledge graph (semantic layer) + Logic graph (logic layer) + Decision graph (decision layer) + an AI expert team, and as the company grows it can connect to real people and IM.

Install one Skill, turn your company into a runnable agent system.

This "company brain" seamlessly switches between three deployment shapes (see references/product_intro.md and references/deployment.yaml):

• A · Skill package (current): runs on the WorkBuddy cloud runtime — subscribe and use, zero ops.
• B · Edge-cloud sync: edge keeps the core, cloud handles the heavy lifting (reference impl: OpenBMB EdgeClaw Box); sensitive data never leaves the domain.
• C · Fully on-prem: the whole spec + runtime self-hosted, fully offline (air-gapped), Palantir-style.

All three shapes share the same vendor-neutral ontology spec — swap the shell and it just works, perfectly self-consistent. It already delivers a readable, writable, closed-loop digital twin of a company, and automates Palantir's FDE (Forward-Deployed Engineer) delivery model into an out-of-the-box AI-FDE — the owner digitizes the company just by chatting.

First step after install (guided instantiation)

New users follow these six steps to describe their company; the skill auto-matches an industry pack, generates an instance, and builds the expert team. Full guidance in GUIDE.md (detailed Phase-5 version); here is the instant entry point:

0. Read the guide: start with GUIDE.md (six-step English guide).
1. Name your company type: one-person company / e-commerce / small business / other.
2. Introduce the company: industry, region (e.g. Tianjin), scale, registration info.
3. Pick departments: the skill ships a 24-department library (12 core defaults + 12 optional) — accept the recommendation or mix freely; you can later rename / set gender / add private departments (see below).
4. Describe business logic: core flows (e.g. order → invoice → bookkeeping → tax filing).
5. Get optimization advice: after building, the skill suggests missing departments / compliance gaps / closed-source upgrade packs (open-core monetization entry).

→ The skill matches the corresponding pack under references/packs/, writes the profile into instance.yaml (separate from the template), and builds the expert team in one click.

Where the company instance lives (persistence · digital-twin source of truth)

The generated instance.yaml is the company's authoritative source of truth (basics, departments, workflows, owner profile, culture, context). It defaults to a stable data directory (separate from the skill code, valid across workspaces):

• Weishu example: C:\Users\li_sh\.workbuddy\company-data\weishu\instance.yaml
• Companion memory log: C:\Users\li_sh\.workbuddy\company-data\weishu\company-memory.md

The digital company (expert team) automatically Reads these two files at the start of every session, collaborating on the same company facts. When the owner confirms a new fact / decision / risk, the steward appends it to the memory log and writes back to the instance when necessary. This way the company state accumulates across sessions instead of being rebuilt empty each time. PII (ID numbers, bank cards, receipt images) is never written in plaintext.

This data directory is the persistent layer of the "digital twin"; when later connecting real systems (knowledge base / IM), the source of truth migrates from here to a database.

File map

text

ontology-agent/
  SKILL.md                  # you are here: entry point
  GUIDE.md                  # six-step user onboarding guide
  references/
    product_intro.md        # product intro copy (for users/channels/investors: 3 modes + digital twin + AI-FDE)
    contact.yaml            # vendor contact info (site/email/phone + enterprise deployment guidance; AI can read & answer)
    deployment.yaml         # three deployment shapes + sync boundaries + runtime options
    portability.md          # portable deployment & integration (Mode B/C landing artifacts, export + integration flow)
    glossary.yaml           # bilingual term standard layer (zh/en reference throughout: 24 dept codes / 18 semantic entities / 8 industry packs / source codes)
    semantic.yaml           # semantic layer template (13 entity types full fields + 28 relation library + 8 global constraints)
    logic.yaml              # logic layer template (references action_specs + stage state machine + industry workflows)
    decision.yaml           # decision layer (13 decision rules + concrete compliance clauses incl. Tianjin policy)
    roles.yaml              # multi-role governance (Phase 3, reserved for enterprises)
    department_actions.yaml # 24 departments / 177 action list (12 core + 12 optional, dedup-merged)
    action_specs.yaml       # 159 action full specs (141 core + 18 optional-dept extensions; desc/inputs/outputs/uses_entities/skill_ref)
    packs/                  # industry packs (Phase 4)
      opc.yaml              # one-person company (first release)
      ecommerce.yaml        # e-commerce
      sme.yaml              # small & micro business
      manufacturing.yaml    # manufacturing factory (12 sub-types in subtypes section)
      public_institution.yaml # public institution / government-affiliated unit (state assets / gov procurement / staffing)
      finance.yaml          # financial institution (AML / suitability / capital regulation, reference non-legal opinion)
      biomed.yaml           # healthcare / biomedicine (genetic resources / GMP / health data)
      edu_research.yaml     # education / research institute (research funding / tech transfer)
    onboarding.md           # progressive proactive onboarding design notes (the digital company owns onboarding)
    customization.md        # user private customization mechanism (rename/add-remove/gender, zero intrusion to skill package)
    monitoring.yaml         # autonomous monitoring signal domains + toggleable automation templates
    proactive.yaml          # spontaneous-behavior four-tier grading + authorization gate
  experts-template/         # 24-department team-building template (12 core + 12 optional)
  assets/avatars/           # avatar pool (24 core × male/female + team icon + generic backups; by gender, distributable)
  connectors/               # DingTalk / WeCom / Feishu connectors (stub, API ports left, no network by default)
  scripts/
    init.py                 # guided interview + pack selection + instance generation + team build
    ops_db.py               # local data layer SQLite (knowledge-base 4 core tables + extension tables, remote master port left)
    onboarding.py           # progressive onboarding engine (reads instance, returns next questions + progress)
    sync_embedded.py        # sync embedded expert package after skill update
    avatars.py              # avatar selection + gender logic + write to expert package
    monitor.py              # autonomous monitoring engine (thin/thick profile parse + signal ingestion)
    proactive.py            # spontaneous-behavior engine (grading + intel briefing)
    export_spec.py          # spec exporter: spec + instance packed into portable bundle
    adapters.py             # integration config generator: vendor-neutral mode list + reference impls + host-platform scaffolds
  examples/                 # generic demo seed (fictional example company, for other users to reference structure, not Weishu-specific)
    demo_company/
  VERSION                   # skill version number (for sync comparison)
  THIRD-PARTY.md / LICENSE / README.md

Connecting real systems (port layer · no network by default)

The skill ships with the framework and ports to connect real systems, but everything defaults to stub state — when any external API credential is missing it only prints a hint and never makes a real network request, guaranteeing install-and-use, no misfires, no leaks.

• Local data layer (scripts/ops_db.py): lands the knowledge-base 4 core tables (todo inbox / department matrix / finance-tax calendar / daily ops report) + extension tables as SQLite, default at company-data/<company>/ops.db, same directory as instance.yaml (digital-twin persistent layer). Remote master integration: see RemoteDatasource in the same file (fill space_id / api_token / endpoint to enable dual-write).
• IM connectors (connectors/): unified DingTalk / WeCom / Feishu interface (send msg / receive events / pull members); fill each platform's CREDENTIALS and complete the TODO to enable.

So you can run the company locally and accumulate operational data right now; once you get DingTalk / WeCom / Feishu or knowledge-base API credentials, fill them in to switch from stub to real integration — no architecture change needed.

Skill update → sync embedded expert package (important)

The expert package (digital company) embeds a copy of this skill. After the skill itself upgrades, the embedded copy does not auto-change. Run the sync script (or let the digital company remind you to run it):

bash

python <skill dir>/scripts/sync_embedded.py <expert package dir>

• At the start of every session the digital company (steward) compares the "embedded copy VERSION" with the "user-level skill VERSION" (~/.workbuddy/skills/ontology-agent/VERSION); if they differ, it proactively prompts you to run the sync command above.
• Sync contents: SKILL.md / GUIDE.md / VERSION / references/ / scripts/ / connectors/ / experts-template/; it never touches your instance.yaml / company-memory.md (that's your data, not the skill itself).

The digital company proactively onboards you (progressive Onboarding)

You don't have to remember "what should I tell the digital company" — the digital company owns onboarding and progressively collects company data by priority: only 1–2 highest-priority gaps at a time, with purpose + reassurance, never a one-time bombardment, never raising security concerns. See references/onboarding.md; at runtime the steward calls scripts/onboarding.py to get the next question and post progress. Generic demo seed in examples/demo_company/.

24-department library & user private customization (no change to skill package integrity)

The skill ships 24 departments (experts-template/department_experts.yaml): 12 core (Lead / Admin / HR / Finance-Tax / Legal / R&D / Content / Channel / SEO / Sales / Data / Strategy, default team) + 12 optional (Customer Service / Product / Design / Ops / Procurement / Quality / IT-Security / PR / Investment-Finance / Customer Success / Cross-border / Risk-Control, activate on demand). This way the general edition still defaults to 12 (preserving the validated experience) while the package is fully stocked for most commercial enterprises.

Users can privately customize any department, and it is always stored only in their own company-data/<company>/instance.yaml, never modifying the skill package itself:

Operation	Command	Touches skill package?
Rename	python scripts/init.py --slug <short> --rename finance "Chief Financial Officer"	No
Set gender (avatar/tone only)	python scripts/init.py --slug <short> --gender sales female	No
Disable dept	python scripts/init.py --slug <short> --disable mkt_seo	No
Add private AI employee	python scripts/init.py --slug <short> --add-dept cso --name "Chief Security Officer" --gender male --expertise "security strategy, compliance governance"	No
List current depts	python scripts/init.py --slug <short> --list-depts	—
Browse 24-dept library	python scripts/init.py --catalog	—

Language flag — --lang en|zh (default en):

Every command above accepts --lang. It switches the kernel data the scripts read, so the same installation serves both audiences:

--lang	Kernel files read	Expert names & terms	Industry pack
en (default)	references/*.yaml, packs/*.yaml	English	English pack
zh	references/*.zh.yaml, packs/*.zh.yaml	Chinese	Chinese pack

Machine keys (action slug, source codes, department codes, entity type names) are byte-identical in both languages, so scripts and instance.yaml behave the same either way.

bash

python scripts/init.py --demo --lang en    # English expert team
python scripts/init.py --demo --lang zh    # Chinese expert team, Chinese kernel

Merge logic in scripts/init.py: resolve_departments(); details in references/customization.md. During progressive onboarding the digital company gently suggests optional departments and explicitly states "these changes live only in your own company data, never modifying the skill package."

Avatars & gender (better experience)

• Avatar pool assets/avatars/ ships: 12 core departments × male/female (24) + team icon team.png + 4 generic backups (2 each gender).
• Each department has a gender_default; after the user overrides with --gender, regenerating/syncing the expert package applies the corresponding avatar. Gender only affects avatar and tone, never department capability.
• Optional 12 departments and custom departments fall back to the generic pool if they have no dedicated avatar; a dedicated avatar can be generated with one command when needed.
• Write the chosen avatar into the expert package: python scripts/avatars.py --slug <short> --pkg <expert package dir>.

Autonomous monitoring & spontaneous-behavior layer (nice-to-have · zero intrusion · Phase 10)

On the premise that the user has not actively assigned work and internal normal operations are unaffected, the digital company additionally "from time to time" digs up information relevant to the enterprise, making the user feel "like a good employee is watching my back." This layer sits above the core framework as an add-on layer, not modifying the 24-department template / 159-action library / decision governance.

• What it monitors: see references/monitoring.yaml — 8 signal domains (policy / industry hotspot / public opinion / business health / competitors / customers / risk / talent), each mapped to a real department + existing action; thin profile (little recorded) runs general domains (policy / industry / opinion / business health), thick profile (more recorded) adds precise domains (competitors / customers / risk / talent).
• Spontaneous logic: see references/proactive.yaml — trigger pipeline (signal → relevance → value score → authorization check → decision → trace) + four tiers: L0 silent (irrelevant, log only) / L1 light reminder (into briefing) / L2 draft awaiting confirmation (default tier) / L3 execute within authorization (only low-risk routine actions).
• Runtime: scripts/monitor.py parses thin/thick profiles and writes signals; scripts/proactive.py grades and aggregates the "intel briefing."
• Guardrails: reuses decision.yaml's PII / compliance red lines and persona-D authorization boundary; any overreach (tax filing / invoicing / major contracts / fund transfer) is always downgraded to L2 draft + await confirmation, never L3.
• On by default: policy inspection (weekly) + ops daily report; the rest (opinion / competitors / customers / risk) one-click enable in the expert center, no disturbance.
• Experience landing: even with little input the user gets "something daily," naturally guided to fill the profile; a thick profile feels like a dedicated executive assistant. The unified deliverable is an "enterprise intel briefing" pushed every 1–3 days.

Three runtime tiers (one spec sold three ways)

• ③ WorkBuddy cloud (single-user SaaS to start)
• ② Edge-cloud sync (enterprise intranet, reference impl: OpenBMB EdgeClaw Box)
• ① Fully local private agent (Palantir-style air-gapped)
• Key: template (ships with this Skill) and instance (generated at user runtime, stored local/private) are strictly separated → data never leaves the domain, the three tiers are universal.

Intellectual property & license

Ontology Agent is a product of Tianjin Weishu Artificial Intelligence Technology Co., Ltd. Its proprietary architecture, integration logic, Chinese localization and domestic deep content, and onboarding engine are original expressions owned by Weishu (© Weishu), may be closed-source and commercial. Underlying referenced open-source components retain their respective authors' copyrights under their licenses — see THIRD-PARTY.md.

Contact (vendor)

This product is made by Tianjin Weishu Artificial Intelligence Technology Co., Ltd. For official site, technical support, or to upgrade to a more professional enterprise deployment, contact Weishu AI through the channels below (see references/contact.yaml, which the AI can read directly and answer user questions):

• Official site: https://ontoagent.net
• Email: [e-mail]
• Phone: [phone]
• Enterprise solutions (custom build / edge-cloud sync / Palantir-style air-gapped private digital-company system): require the enterprise to provide its own local compute server and related hardware; contact Weishu AI via the above channels for technical support.

Industry packs (built-in company-type templates)

The skill matches a pack under references/packs/ by company type, injecting industry-specific entities / department emphasis / flavor / red lines into the instance, zero intrusion to the general 24-department library.

Pack	For	Department strategy
opc.yaml	one-person company (first release)	12 core
ecommerce.yaml	e-commerce	12 core + e-commerce flavor
sme.yaml	small & micro	12 core + SME flavor
manufacturing.yaml	manufacturing factory	12 core + ops/procurement/quality + sub-types
public_institution.yaml	public institution / government-affiliated	12 core + ops/procurement/quality + state-asset/internal-audit/party-gov (private customization)
finance.yaml	financial institution (bank/securities/insurance/fund/trust/payment)	12 core + risk-control/IT-security/investment + compliance AML (private customization), red lines are reference non-legal opinions
biomed.yaml	healthcare / biomedicine (hospital/pharma/device/genetics)	12 core + quality/IT-security/customer-service + medical/registration/pharmacovigilance (private customization)
edu_research.yaml	education / research institute (school/institute/new R&D)	12 core + ops/quality/investment + academic/Research-management/tech-transfer (private customization)

Manufacturing factory pack (manufacturing.yaml) uses "one main pack + built-in subtypes section" to cover sub-sectors without a separate file each (avoids bloat):

• Common: factory/line/equipment/process/work-order/yield/capacity/material/BOM/supplier entities; produce→QC→deliver core loop; safety/environment/quality-system (ISO9001·IATF16949·GMP·CCC)/special-equipment/fire-protection red lines (decision.yaml manufacturing_safety); manufacturing-specific monitoring signals (downtime/OEE, yield fluctuation, material break, safety-env incident, energy, recall — see monitoring.yaml manufacturing domain).
• 12 sub-types (auto-matched by match_keywords, only differences filled): apparel-textile / food-beverage / electronics / machinery (heavy) / chemical / building-materials / pharma-bio / FMCG / automotive-parts / metal-smelting / packaging-printing / plastic-toys.
• Manufacturing "heart" departments (production/manufacturing, equipment-engineering, process-engineering) are not in the 24-department general library; suggested via suggested_custom_departments for the user to privately add, no change to skill package integrity.

Portable deployment & integration (Mode B / Mode C landing)

The spec layer is vendor-neutral data; three modes = same spec + different runtime shells. Swap the shell and it works, the business brain is not lost, no rewrite.

• Export spec: python scripts/export_spec.py --bundle all --slug <short> --zip → generates self-contained exports/<short>_ontology_bundle/ (all references/ + instance + manifest.json contract). --bundle source exports only the blank spec template, --bundle instance only the digital twin.
• Integrate runtime: python scripts/adapters.py --bundle exports/<short>_ontology_bundle --pattern edge_cloud (vendor-neutral mode list, main output)
  • Reference impls (pick any): --target edgeclaw (OpenBMB EdgeClaw Box) / --target joyagent (JD JoyAgent-JDGenie) / --target both
  • Mode B edge-cloud sync (ref impl: OpenBMB EdgeClaw Box): produces edgeclaw_manifest.yaml, sync_boundary keeps sensitive on edge, external brain to cloud, auto splits edge/cloud by department sensitivity (lead/admin/hr/ finance/legal stay on edge).
  • Mode C fully on-prem (ref impl: JD JoyAgent-JDGenie): produces joyagent_manifest.yaml + joyagent.docker-compose.yml (deploy template hint), fully offline, local model, spec zero-rewrite, ERP/MES write-back hooks.
• Host-platform "wrap scaffold" (answering "is it WorkBuddy-only?"): python scripts/adapters.py --slug <short> --target coze|dify|claude|yuanqi|qianfan|workbuddy → ports the same vendor-neutral spec to competitor platforms, spec zero-rewrite.
• Both scripts read-only references / instance, never modify the core; the integration manifest only references the spec (./references, ./instance), zero-rewrite — this is the ground of "perfectly self-consistent access." Full flow in references/portability.md.

Term convention (bilingual term standard layer)

All yaml machine keys are English, user-facing text is Chinese; the mapping lives in references/glossary.yaml. Register new terms in the glossary before use.

Standard layer scope: glossary.yaml already covers — core four layers / atomic concepts / packaging-instantiation / governance-security / 8 industry packs / three runtime tiers / connectors / source codes / 24 department codes zh-en / 18 semantic entities zh-en / reference-library sources. The standard layer is injected into the expert system_prompt as a "term standard layer · zh-en mapping" appendix (see scripts/init.py: glossary_block()), ensuring every AI employee uniformly uses Chinese terms while recognizing internal English identifiers — the standard layer is actually used, not left idle. IP registration note: user-facing terms carry no "original/borrowed" tag; the source-code section only records action provenance (OPB/Frog/slav/ReS/sober/easy are borrowed, WSHU is Weishu's own), registered in a dedicated attachment under the "cite + lightweight modification + no exclusive claim" discipline.