docs(mvp): add harness design and progressive guides
This commit is contained in:
zhuyongxin
@@ -40,11 +40,28 @@
|
||||
|
||||
---
|
||||
|
||||
## Harness
|
||||
|
||||
| 文档 | 内容 |
|
||||
|---|---|
|
||||
| [harness/README.md](harness/README.md) | **从这里开始**：用一次诊断请求理解 Harness，不要求先掌握组件和状态 |
|
||||
| [harness/components/README.md](harness/components/README.md) | **组件渐进式导读**：沿一次请求分四步理解运行控制、Agent 收敛、Tool 事实边界和验证发布 |
|
||||
| [harness/CONTEXT.md](harness/CONTEXT.md) | Harness 统一术语、命名规则、状态维度与已知命名债务 |
|
||||
| [harness/Harness生命周期与状态.md](harness/Harness生命周期与状态.md) | Run、Tool、Progress、Guard、Release、SSE 和持久化生命周期及状态映射 |
|
||||
| [harness/Harness设计-非确定性Agent的确定性控制边界.md](harness/Harness设计-非确定性Agent的确定性控制边界.md) | Harness 根问题、设计不变量、方案取舍、关键决策、代价与真实问题反推 |
|
||||
| [harness/Harness组件全景-职责-设计原因与边界.md](harness/Harness组件全景-职责-设计原因与边界.md) | 当前 Harness 10 个职责域、全部生产类型、设计原因、作用与边界 |
|
||||
| [harness/Harness-Tool双视图-从原始结果到可验证证据.md](harness/Harness-Tool双视图-从原始结果到可验证证据.md) | canonical truth、Control View、Agent Observation 与 metadata audit 的数据边界 |
|
||||
| [harness/Harness证据安全链-从引用真实到结论可发布.md](harness/Harness证据安全链-从引用真实到结论可发布.md) | EvidenceGuard、EvidenceRepair、SemanticGuard 和唯一 Release Policy |
|
||||
| [harness/Harness信息增益停止-让无证据诊断正常收敛.md](harness/Harness信息增益停止-让无证据诊断正常收敛.md) | GAINED/NO_GAIN、重复检测、协议停止、ProgressSnapshot 与过程型 Fallback |
|
||||
|
||||
---
|
||||
|
||||
## 诊断 / E2E
|
||||
|
||||
| 文档 | 内容 |
|
||||
|---|---|
|
||||
| [diagnosis/一次诊断全流程-E2E导读.md](diagnosis/一次诊断全流程-E2E导读.md) | SUCCESS 全流程：阶段、token、timeline、字段 |
|
||||
| [diagnosis/Session-Run-Trace隔离-从串线到可回放.md](diagnosis/Session-Run-Trace隔离-从串线到可回放.md) | 多轮串线问题、身份拆分、协议迁移与 exact-run E2E |
|
||||
|
||||
架构对照：
|
||||
|
||||
|
||||
Reference in New Issue
Block a user