g0-gate

G0 Gate — AI Employee Governance: Session-Opening Context Assembly

AI 员工治理 · 开工闸门：让 AI Agent 每次开局都装上”正确档位、验证已加载”的上下文——不靠人粘贴，不装睁眼瞎。

G0 Gate is a session-opening protocol for AI agents. It sits before G1–G4 (task-discipline gates) and owns one thing only: context assembly quality at session start. The system carries the rigor, so humans don’t have to paste protocol text, guess the right tier, or chase expired documents.

G0 (context assembly) → G1 (grill-me) → G2/G3/G4 (task discipline)

Why this matters (the AI employee governance problem)

• Agents start blind: without a gate, an agent opens a session with missing/wrong/expired context — and acts confidently on it.
• Human habit assets don’t scale: relying on someone to paste the right protocol is fine for one operator, impossible for customers running AI employees.
• The fix: a gate the system runs — auto tier selection, auto load, self-verified.

If you’re building AI employees / AI agent teams / multi-agent systems and care about AI employee governance (discipline, accountability, reproducible behavior), G0 is the entry gate you’re missing.

Core ideas

Idea	What it does
Tiered manifests	lite / standard / heavy — the system picks by task type + window budget, humans don’t memorize
Window budget guard	Constraint = context window, not tokens; low budget → auto-degrade + report what was skipped
Self-check loop	Every source returns ok/stale/missing/unreachable → aggregate context_trust: full/partial/none; never pretend to have read something
Domain routing	Load only the protocol sections + ≤3 skills for the task domain; no full scans
Versioned registry	g0_registry.json: protocol edit → bump version → next session effective; the registry file is the sync channel
Enforcement (A/B/C)	Harness hard-block (A) + machine auto-score (B) + embodied assistant voice (C) — skipping becomes impossible, not merely discouraged

Quick start（30 秒跑通）

先看它跑起来的样子（30 秒数字人演示）：

# 1. 看 G0 跑起来的样子（无依赖，纯 Python）
python examples/run_g0_demo.py

# 2. 跑 8 个验证场景（全 PASS = 协议实现正确）
python examples/verify_gate.py --all

repo/
├── spec/
│   ├── G0-gate-spec-v2.md   # 🔴 现行规格 v2.0（BSL 1.1，含档位算法/manifest v2/自检公式）
│   ├── wire-protocol.md     # /gate/init + /gate/run + /gate/status 端点契约
│   └── verification.md      # 8 场景验证方法学
├── examples/
│   ├── g0_registry.example.json   # manifest v2 schema
│   ├── run_g0_demo.py             # 30 秒模拟演示
│   └── verify_gate.py             # 8 场景自动化断言
├── docs/
│   ├── migration-guide.md   # v1.0(MIT) → v2.0(BSL) 迁移说明（v1 已过时）
│   └── SKILL.md             # Layer A agent-side prototype
└── LICENSE                  # BSL 1.1（商用需授权）

1. Read the spec → spec/G0-gate-spec-v2.md（v2.0 现行；v1 见 docs/migration-guide.md 已过时）

2. Add a session-init hook to your agent runtime that calls G0.run() on every new session:

curl -X POST http://127.0.0.1:PORT/api/v1/session/start \
  -H "Content-Type: application/json" \
  -d '{"agent_id":"my-agent","project":"my-project","g0":true,"g0_tier":"heavy","window_budget":55}'

3. Point the registry at your passport/daily/records sources (see examples/g0_registry.example.json).

4. Wire enforcement: harness blocks final output until gate = pass; machine scores gate-skipping; the assistant (if embodied) narrates the “facts I can check vs decisions I need from you” split.

Example registry

{
  "version": "1.0.0",
  "defaults": {
    "internal_aliases": ["steward", "lead", "operator"],
    "tier_policy": "auto",
    "budget_threshold": 40,
    "skill_cap": 3
  },
  "tiers": {
    "lite":     { "sources": ["passport:50", "records:5"] },
    "standard": { "sources": ["passport:full", "daily:1", "records:10", "framework:grep"] },
    "heavy":    { "sources": ["passport:full", "daily:3", "records:20", "framework:full", "gates:all"] }
  },
  "projects": {
    "company-ai": { "tier_policy": "auto", "domain": "engineering" }
  }
}

Acceptance criteria (from SPEC §7)

• New session → no human pasting; agent auto-pulls correct-tier manifest (field-tested)
• Any source unreachable → marked untrusted, does not block
• Window budget low → auto-degrade + report skipped items
• Protocol edit → registry bump → next session effective

Status

Production-tested in a multi-agent studio (internal + customer agents). Version 1.0.

License

BSL 1.1 — see LICENSE.

This site is open source. Improve this page.