Agent Sop

agent-sop

A human-gated, capability-slotted collaboration workflow (SOP) for coding agents.

[中文说明 →](README.zh.md)

The goal is not to make agents more autonomous — it is to pick the right process weight for each risk level, so that work stays traceable, reviewable, and signable without small tasks drowning in ceremony.

High-risk work follows the full auditable sequence:

research → plan → human approval → implementation → independent review
        → verification → signoff packet → human signoff → deploy

Low-risk work runs a deliberately lighter path. Tiers and boundaries are defined in `SKILL.md` §1.1.

How it works

flowchart TD
  subgraph DISC["🔍 Discovery"]
    direction TB
    A([Task start]) --> B[discover-roster.py --probe]
    B --> C[Slots by condition, not name]
  end

  subgraph TEAM["💜 Team"]
    direction TB
    C --> D[sole-writer: strongest + write]
    C --> E[reviewer: strongest, other vendor]
    C --> F[scouts: cheapest fast tier]
    D & E & F --> G[Roster proposal]
  end

  subgraph PIPE["⚙️ Pipeline"]
    direction TB
    G --> H[[HUMAN: approve / veto roster]]
    H -->|veto| C
    H -->|approve| I[Plan]
    I --> J[[HUMAN: approve plan]]
    J --> K[Sole-writer implements]
    K --> L[Cross-vendor independent review]
    L --> M[Verification]
    M --> N[Signoff packet]
    N --> O[[HUMAN: signoff]]
  end

  subgraph SHIP["🚀 Ship"]
    direction TB
    O --> P[Deploy]
    P --> Q([Agents never self-serve publish or deploy])
  end

  classDef startEnd fill:#0EA5E9,stroke:#0369A1,stroke-width:2px,color:#fff,rx:12,ry:12
  classDef discover fill:#38BDF8,stroke:#0284C7,stroke-width:2px,color:#0C4A6E
  classDef agent fill:#34D399,stroke:#059669,stroke-width:2px,color:#064E3B
  classDef proposal fill:#86EFAC,stroke:#16A34A,stroke-width:2px,color:#14532D
  classDef plan fill:#6EE7B7,stroke:#0D9488,stroke-width:2px,color:#134E4A
  classDef work fill:#4ADE80,stroke:#15803D,stroke-width:2px,color:#14532D
  classDef re