Session Workflow Beta
A Session is the second core data model in Forge. Every product discussion is captured in a Session, holding the full timeline of conversation, decisions, and Agent outputs.
How Sessions relate to the Ontology
Session Model
Session {
id // sess_...
title // "添加成本告警"
participants // [人类, PM Agent, Architect Agent, Code Agent]
ontology // 关联的 Ontology 节点列表
turns // 对话轮次
artifacts // 产出物:PRD、ADR、PR 链接
status // active / paused / shipped
}Three Kinds of Agents
From a vague requirement to a merged PR, three kinds of Agents collaborate in a relay.
Breaks down vague requirements into user stories, acceptance criteria, and priorities.
Architecture modeling, technology selection, and risk identification; writes into the Ontology Architecture subtree.
Writes code, runs tests, and opens a PR based on the outputs of the previous two.
The Full Flow from 0 to PR
flowchart TD
H["人类: 想做一个成本超 50 元每日告警"]
H --> PM["PM Agent: 拆为 3 个 UserStory"]
PM --> US1["US-1 实时聚合每日成本"]
PM --> US2["US-2 超阈值触发通知"]
PM --> US3["US-3 通知模板配置"]
US1 --> AR["Architect Agent: 决定技术方案"]
US2 --> AR
US3 --> AR
AR --> A1["数据源: Lumen CostTracker API"]
AR --> A2["调度: 每 1 分钟 cron"]
AR --> A3["通知渠道: Webhook + 邮件(复用 Lurus Platform)"]
AR --> A4["写入 Ontology: Architecture/cost-alert.md"]
A1 --> CD["Code Agent: 代码实现"]
A2 --> CD
A3 --> CD
A4 --> CD
CD --> C1["生成 Go handler + test"]
CD --> C2["生成 DB migration"]
CD --> C3["开 PR: feat(cost-alert): add daily threshold"]
C3 --> C4["等待人类 review & merge"]WAL Decision Traceability
Built on the WAL of the Kova engine, every step of conversation and decision is persisted. At any time you can:
Trace back "why NATS was chosen over Redis Streams"
Pinpoint "which Session last wrote to Architecture/auth.md"
Replay an entire Session for a retrospective