dong-ai 0.3.1

Dong AI Company — 多智能体 AI 公司框架

Dong AI Company

Cross-project memory and organizational governance for AI agents.

pip install dong-ai
dong demo          # see it in action — no API key, no network calls
dong setup         # configure providers
dong make "a quarterly market analysis"

---

See it in action (no API key required)

dong demo creates two fake projects in an isolated local SQLite database and shows how graph memory enables cross-project search, symbol drill-down, dependency impact analysis, and session resume. Zero config, zero API key, zero network calls.

---

The Problem

Every AI agent works in a vacuum. Close a conversation, lose the context. Start a new project, the model has no memory of last week's architecture decisions. Long-running projects degrade into incoherence as the window fills with noise.

Dong AI solves this by adding two missing layers to any agent:

1. Cross-project indexed memory — every symbol, decision, and dependency is stored in a queryable graph that persists across sessions. The model fetches what it needs instead of scrolling through degraded history.
2. Organizational governance — red/blue debate before design decisions, dynamic worker recruitment per task, board review with quality gates, and automated post-project debrief that extracts lessons for future work.

The result: agents that remember what they built, why they built it that way, and improve with every project.

---

Quick Start

pip install dong-ai        # zero external AI deps
dong setup                  # detect hardware, configure providers
dong make "a CLI tool for CSV parsing"

No API key required to start — dong check shows available options.

All Commands

Command	Use
dong make "request"	Self-directed execution for any domain
dong run "request"	Full governance pipeline (debate → workers → review)
dong quick "task"	Lightweight mode, no pipeline overhead
dong analyze <path> 'question'	Read and analyze any source file
dong edit <path> 'instruction'	Edit file with diff preview
dong debug	CI failure root cause analysis
dong company start --domain "..." --duration 8h	Background company runtime
dong company status	View running company state
dong company knowledge	Organizational metacognition map
dong company review	Decision audit trail
dong graph list	List indexed graph projects
dong graph view <id>	Drill into a project's symbols and dependencies
dong detect	Full system status
dong serve	OpenAI-compatible API server
dong update	Upgrade from PyPI

---

How It Works

Any Agent (CLI / Hermes / Claude Code / Cursor / Copilot)
  │  shell out: dong <command>
  │
  ▼
Dong AI Engine
  │
  ├── Column Memory          5-slot context management (C0-C4) with unloading
  ├── Graph Memory           Persistent symbol/dependency/decision index
  ├── Experience Engine      Post-project debrief → skill extraction → future recall
  ├── SafetyGovernor         Confidence scoring, risk budgets, confirmation gates
  ├── Metacognition          Knowledge map, strategy evolution, improvement tracking
  └── Domain Runtimes        7x24 background operation for monitoring tasks

Column Memory

Not "infinite context" — precise context. Five named columns with independent token budgets:

Column	Content	Reason to keep
C0	Project goal, constraints, user	Always available
C1	Symbol signatures, API contracts	Reference surface
C2	Dependency maps, architecture	Navigation
C3	Active decisions, rationale	Traceability
C4	Historical context, past output	Full when room

When the budget is tight, C4 is unloaded first. The model always has the current goal, available symbols, and recent decisions — without wading through 50K of conversation history.

Graph Memory

Every project run indexes its symbols, dependencies, and decisions into an SQLite graph. Three retrieval methods:

• Keyword match — exact signature lookup
• Semantic search — embedding cosine similarity
• Graph traversal — dependency chain walking

Impact analysis on every query: load_config → 2 callers, risk score: 50%.

Projects don't exist in isolation. Symbols from one project can be referenced by another. Cross-project queries work across your entire work history.

Governance Pipeline

User Request
  │
  ├── CEO — type detection, pipeline generation, worker recruitment
  ├── Design — red/blue team debate, requirements extraction, coverage checklist
  ├── Execute — dynamic workers, parallel execution, self-healing (×3), cross-review
  ├── Board — phase scoring (1-10), minimum gate 6.0, requirement audit
  ├── Debrief — lesson extraction → Experience Engine
  └── Report — full evidence trail

No hardcoded roles. Each task generates its own specialist team via LLM — software projects get architects and engineers, novels get world-builders and writers, audits get security analysts.

---

Capabilities

Category	Description
Project execution	Three modes: make (self-directed research→execute), run (full governance), quick (lightweight)
Code workflow	analyze (code Q&A), edit (diff-preview edits), debug (CI root cause from GitHub Actions)
Company runtime	7x24 background domains, configurable duration, hourly health checks, daily reports
Metacognition	Knowledge map of domain expertise, learning strategy effectiveness, knowledge gaps
Governance	Confidence scoring, risk budgets, confirmation gates, decision audit trail
Context	5-column memory management, cross-project graph memory, session recovery
Plugins	MCP ecosystem integration, plugin registry
Models	20+ providers with automatic failover, local GGUF support, custom provider config
API	OpenAI-compatible server, webhook subscriptions, scheduled cron tasks

---

Why Not Just Use a Bigger Context Window?

Larger windows don't solve the structural problem — the model still has to search through noise to find signal, and the window is reset when the conversation ends.

Dong AI's approach is orthogonal: index instead of remember. A 1K-2K precision query replaces 50K of degraded context. Cross-session persistence replaces per-session ephemeral state. This works regardless of window size — 64K, 128K, or 1M.

---

Integration

Dong AI is designed to be called from any agent that can run shell commands.

Agent	Method
Hermes Agent	cp SKILL.md ~/.hermes/skills/dong-ai-company/
Claude Code	cp CLAUDE.md ~/.claude/claude.md
Cursor	cp .cursorrules .cursorrules
GitHub Copilot	.github/copilot-instructions.md
Any	dong <command> from shell

Adapter files are maintained at Dong04-123/Dong-AI-skill.

---

Testing

pip install pytest
pytest tests/
# 281 tests, ~2s, zero external deps, no network calls, no API keys

---

Project Status

Active development. v0.1.x — core pipeline, column memory, experience engine, governance, company runtime, and multi-agent adapters are functional. Breaking changes possible until v1.0.

---

License

MIT — free for personal, research, and commercial use. Attribution required.

---

██████╗  ██████╗ ███╗   ██╗ ██████╗     █████╗ ██╗
██╔══██╗██╔═══██╗████╗  ██║██╔════╝    ██╔══██╗██║
██║  ██║██║   ██║██╔██╗ ██║██║         ███████║██║
██║  ██║██║   ██║██║╚██╗██║██║         ██╔══██║██║
██████╔╝╚██████╔╝██║ ╚████║╚██████╗    ██║  ██║██║
╚═════╝  ╚═════╝ ╚═╝  ╚═══╝ ╚═════╝    ╚═╝  ╚═╝╚═╝

Engine · PyPI · Adapters · Hermes Agent