Evidence：用真实观察检查完成性

用 App 定义的证据语义检查需求，避免把模型的“完成了”当成真实回执。

Checklist 按 requirement.kind 选择已登记的 isSatisfied。isSatisfied 检查一项，isComplete 检查全部，missing 返回未满足需求。未知 kind 拒绝；predicate 必须同步返回 boolean。Kit 不推断“文件已交付”或“事务已提交”，App 必须用真实结果定义。

EvidenceProjection 按 sourceIds 缩小带来源的观察。未选择或空数组时保留完整证据，未知工具也保持原观察；这与按权限过滤不是一回事，权限应在来源读取前完成。默认工具选择读取 toolId，再兼容观察本身提供的 tool；App 可显式提供 toolId()。

由简到繁的同业务示例

检查真实资料是否已读取

入门 · Node.js

TypeScript

import { Evidence } from '@jangts/agent-core/kits/evidence';
type Observation = { sourceId: string; read: boolean };
type Requirement = { id: string; kind: 'source-read'; sourceId: string };
const checklist = Evidence.defineEvidenceChecklist<Observation, Requirement>([
  { kind: 'source-read', description: '资料读取已成功', isSatisfied: (seen, need) => seen.read && seen.sourceId === need.sourceId },
]);
console.log(checklist.isComplete({ observations: [{ sourceId: 'manual', read: true }], requirements: [{ id: 'need-manual', kind: 'source-read', sourceId: 'manual' }] }));

JavaScript

import { Evidence } from '@jangts/agent-core/kits/evidence';
const checklist = Evidence.defineEvidenceChecklist([
    { kind: 'source-read', description: '资料读取已成功', isSatisfied: (seen, need) => seen.read && seen.sourceId === need.sourceId },
]);
console.log(checklist.isComplete({ observations: [{ sourceId: 'manual', read: true }], requirements: [{ id: 'need-manual', kind: 'source-read', sourceId: 'manual' }] }));

相同清单返回缺口，再限定审查所用来源

配置 · Node.js

TypeScript

import { Evidence } from '@jangts/agent-core/kits/evidence';
type Observation = { sourceId: string; read: boolean };
type Requirement = { id: string; kind: 'source-read'; sourceId: string };
const checklist = Evidence.defineEvidenceChecklist<Observation, Requirement>([
  { kind: 'source-read', description: '资料读取已成功', isSatisfied: (seen, need) => seen.read && seen.sourceId === need.sourceId },
]);
console.log(checklist.missing({ observations: [], requirements: [{ id: 'need-manual', kind: 'source-read', sourceId: 'manual' }] }));
type ToolObservation = { toolId: string; sources: { id: string; text: string }[] };
const projection = Evidence.defineEvidenceProjection<ToolObservation, { id: string; text: string }>({
  sourceId: source => source.id,
  providers: { search: { sources: row => row.sources, withSources: (row, sources) => ({ ...row, sources }) } },
});
console.log(projection.project({ observations: [{ toolId: 'search', sources: [{ id: 'manual', text: '使用指南' }, { id: 'faq', text: '常见问题' }] }], sourceIds: ['manual'] }));

JavaScript

import { Evidence } from '@jangts/agent-core/kits/evidence';
const checklist = Evidence.defineEvidenceChecklist([
    { kind: 'source-read', description: '资料读取已成功', isSatisfied: (seen, need) => seen.read && seen.sourceId === need.sourceId },
]);
console.log(checklist.missing({ observations: [], requirements: [{ id: 'need-manual', kind: 'source-read', sourceId: 'manual' }] }));
const projection = Evidence.defineEvidenceProjection({
    sourceId: source => source.id,
    providers: { search: { sources: row => row.sources, withSources: (row, sources) => ({ ...row, sources }) } },
});
console.log(projection.project({ observations: [{ toolId: 'search', sources: [{ id: 'manual', text: '使用指南' }, { id: 'faq', text: '常见问题' }] }], sourceIds: ['manual'] }));

策略与实现

• 完成性与证据投影规则