AI Company Framework

Company Is a Word. — Say one sentence, get a working multi-agent company.

✨ What it is — in 30 seconds

A DeepSeek Harness dsh.bundle that turns one sentence into a company of role-specialized agents (writer, editor, developer, QA, sales, customer service…) that designs, schedules, quality-checks and delivers — while you stay in charge of every major decision.

你说一句话 → 老板澄清(一次一问) → 实时盘点模型/分岗位路由 → 建团队(AgentTeams)
  → 员工入职自写技能 → 派首单 → 质检闭环 → 交付包(含假设清单) → 你要不要改？一句话定向返工

Architecture

⚡ 30-second quick start

# 1. Install (packed tarball — not a source directory)
dsh plugin --profile web add .\ai-company-framework-0.3.10.tgz

# 2. Restart DSH Desktop, open a NEW session, then:
/company
# 3. Say what you want:
#    "I want an ecommerce content company — first order: a Xiaohongshu post + customer-service scripts."

That's it. The boss agent takes it from there.

AgentTeams activity	Employee sidebar	Company created

🎯 Why it stands out

Advantage	What you get
Zero impact on daily chat	Company mode is off by default — `/company` to enable, `/no-company` to disable. Your normal conversations are untouched until you opt in.
Customer memory that follows the customer	SQLite per-customer profiles (id-isolated) + company/general memory. The same customer across private chat and multiple groups is automatically merged into one profile, shared by every client-facing role — nobody asks the customer to repeat themselves.
Agents push to you — you don't poll	After "let me check", the agent reads its own output and proactively sends the result. No polling, no "did you get it?".
Humans own big decisions	Pricing exceptions, refunds, commitments, legal/PR → a human decision gate pushes the choice to your phone via Feishu; nothing lands until you approve.
One bot per employee	Create a Feishu bot for the boss and any employee with one scan (registerApp) — perfect for remote work and client-facing teams.
Hardened conversation isolation	Sidebar state is session-scoped (fail-closed), model reconfiguration is ownership-checked (403), customer archives are id-isolated. No cross-company leaks.
Open-source safe	The package ships zero user data — databases, logs and credentials (DPAPI-encrypted) are created locally at first use.
Memory mode per role	`三层` (three-layer, client-facing) or `单层` (single, internal) — declared in each role skeleton.

🏭 Use cases

Ecommerce content studio — research → copywriting → QA → rework → delivery, batched into one deliverable.
Game studio — designer plans, writer writes store copy, QA gates, customer service handles clients with full customer memory.
Customer support & sales — per-customer archives + cross-group merge + autonomous push + human decision gate.
Remote team — every employee can have their own Feishu bot; you steer from your phone.
Agency / outsourcing — one-sentence onboarding per client, isolated companies per session.

📦 What's in the box (v0.3.10)

15 flat Skills (all gated by /company):

Skill	Role
`company-boss`	Boss — architecture, scheduling, QA loop, delivery, wisdom
`company-pipeline`	8-stage automation pipeline (build/feishu/schedule/QC/deliver/hire)
`company-role-template`	Role-skeleton template (self-written skills by employees)
`company-customer-memory`	Generic three-layer memory for any client-facing role
`role-writer / role-editor / role-coder / role-researcher / role-qa / role-ops / role-finance / role-hr / role-data / role-translator / role-customer-service`	Prebuilt role skills

Plus:

7 workflow templates (core/templates/).
AgentTeams runtime + web activity panel (bundled dependency, own Cordis row).
Employee sidebar (host + web UI): lossless per-member model routing, live session/tool view, messaging, session isolation, activity-panel restore.
Feishu bot bridge (host + web UI): official registerApp one-scan onboarding, DPAPI credentials, boss + per-employee bots, private/group routing, two-way logs, boss auto-reply, autonomous push, human decision gate, external-contact guide, full-coverage scopes + extraScopes.
Customer memory: core/scripts/customer-memory.mjs (Node built-in SQLite, zero deps) — per-customer profiles + message log, three-layer memory, cross-group/private merge, one-shot full customer pull.
Company mode switch: /company / /no-company commands + company_mode tool.
No RAG, no vector database, no telemetry.

🏗️ Architecture & boundaries

DeepSeek Harness (models, tools, sessions, Skills, sandbox, web)
  └─ AI Company Framework bundle
       ├─ 15 Skills provider (package-local, no user-dir copies)
       ├─ AgentTeams runtime + activity panel (bundled dep)
       ├─ Employee sidebar host + UI (session-isolated)
       ├─ Feishu bridge host + UI (multi-bot, DPAPI)
       └─ Customer memory (SQLite, per-customer)
  └─ Human gates: model API key · Feishu authorization · major decisions · workspace

Models/tools/sessions/Skills/Bundle runtime come from DeepSeek Harness; the rest is implemented by this Bundle and activates on install.
Honest-state rule: anything not installed/authorized/online is shown as such with onboarding guidance — never falsely reported ready.

🛡️ Security

Zero user data in the package (verified: no .db/.jsonl in the tarball).
Feishu App Secrets: Windows DPAPI, CurrentUser scope, local only; connected only when a real WebSocket is up.
Customer archives: per-id isolation; sidebar: per-session fail-closed + 403 ownership checks.
No postinstall hooks; install/uninstall managed by dsh plugin.

✅ Verification

Verified on DSH Desktop @deepseek-ai/dsh 0.1.0-rc.8 (Windows): dual Cordis rows, 15 gated Skills, AgentTeams full lifecycle, sidebar isolation, Feishu end-to-end (boss/staff/group/auto-reply/push/decision gate), customer-memory merge & isolation, release tarball integrity. See RELEASE_NOTES and Wiki.

📚 Documentation

Wiki · Quick start · Architecture · Bundle & extension guide
FAQ · Troubleshooting · Feishu onboarding SOP

🤝 Contributing

PRs welcome — see CONTRIBUTING and CODE_OF_CONDUCT.

Notices

Incorporated components (employee sidebar, Feishu bridge) and the supply-chain boundary are documented in NOTICE.md.

ai-company-framework

Installer

README