Files
AgentLab/frontend/src/dev/fixtures/challenge.tsx
T
ThearchyHelios 4c3ba8bc1b 期 4 复核修正(第 1 轮,前端):核查缺口分「核查未通过」「核查未完成」两组并补上证据页、印章的检查,写目录预告推断时算上被挡下的同项指标,反方审稿概况的说法与排布改准
- 出具缺口:双盲复算的「指标无法复算」是没核成,不再和两路不一致、结论被推翻一起叫「核查未通过」——splitIssuanceGaps 分成 failed「核查未通过」、unfinished「核查未完成」两组,时间线那一行、证据页、成果节点印章的悬停说明都按先未通过、再未完成、最后校验不完整排。check-decode 按服务端 GAP_* 和 _refuted_gap 的原话加出具契约的前缀拼出缺口再分组,服务端改了说法先变红;新预览用例 verify-gaps 画真的 EvidencePane 和成果节点卡片,check-verify 断言两组各列各的、横幅的「校验未全部完成」里没有它们、印章说明的先后,以及 390px 深浅色不溢出、可读(以前去掉这两处的实现所有检查照样全绿)
- 时间线:无法复算那一行只说「已记为缺口」,不再一律说「按缺口计入出具」(探索运行不出具,接在侧枝上的缺口进不了出具);dual_check_catalog 带值的写法补上检查,和审批卡「会记为」用同一个函数、一字不差
- 双盲复算审批卡:工件里没有合并预告时(老工件、组合超上限、还没取到),被目录挡下、带着写到哪的同项指标也算进来——业务日期、去重规则说法不同就说都不记(服务端先归拢、轮不到对目录),有效记录条件同项上有被挡下的只说可能记、也可能不记;推出来的条目标 data-inferred,结尾换成「实际写入以时间线上的记录为准」。check-verify 用新夹具 time_rejected 断言销售额不再说「会记为」
- 反方审稿审批卡:「没能」改「未能」,提示不再说「没有真正查询数据」(触顶的查过),原因补上报告未引用数据库查询,时间线的下一步不再建议调高上限(这一条不含触顶的);「N 句未能反查」挪到降档说明后面,不再插在「出具会降档」标记和它的说明中间,句数不超过「无法判断」;不在出具路径上的那句写明是反方审稿节点(审批卡上的「本节点」读着像审批节点)。ChallengeSummary 收 runClass,给了就照它、不另取;check-challenge 的夹具改成和服务端约定一致(未能反查 2 句、无法判断 3 句),另断言给了运行类别时另取失败也照样写降档、不发请求
2026-10-06 09:31:28 +08:00

255 lines
14 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
import { useState, type JSX } from 'react'
import { ReactFlow, ReactFlowProvider } from '@xyflow/react'
import { Inspector } from '../../canvas/Inspector'
import { NodeCard } from '../../canvas/NodeCard'
import { ApprovalCard } from '../../run/RunPanel'
import { ChallengeSummary } from '../../run/ChallengeSummary'
import { useCatalog } from '../../store/catalog'
import { useStudio } from '../../store/studio'
import type { Approval, ApprovalChallenge, ApprovalReport, Provider, Workflow } from '../../types'
import fx from './challenge.json'
/**
* 反方审稿的预览用例,scripts/check-challenge.mjs 按它们检查。夹具在 challenge.json(检查脚本读同一份):理由文档是后端
* run_verify_agent + compose_doc 真跑出来的,库、表名、地名、数字都是虚构的。
*
* 审批卡(payload.reports[0].challenge):
* - ?syn=challenge-approval:正式运行,没写 on_refuted(按降档),四种判断各一句,有被推翻的 →「出具会降档」
* - ?syn=challenge-approval-flag:正式运行,只标注,不写降档
* - ?syn=challenge-approval-explore:探索运行,不说任何和出具有关的话
* - ?syn=challenge-approval-none:审的报告没有经过反方审稿,什么都不画
* - ?syn=challenge-approval-many:十二句,条目只列前十句,「显示全部」从工件里补齐
* - ?syn=challenge-approval-broken:工件取不到(检查脚本先答 404,再改成能取到,测重试)
* - ?syn=challenge-approval-dense:同第一个,摆在 360px 的右栏里
* - ?syn=challenge-approval-clear / -clear-flag:正式运行,没有被推翻的结论(降档 / 只标注),两种都不写降档、不写只标注
* - ?syn=challenge-approval-explore-flag:探索运行配只标注,同样不说任何和出具有关的话
* - ?syn=challenge-approval-runfail:老服务端的审批(不带 run_class),运行类别也取不到(检查脚本先答 500):不说任何
* 和出具有关的话。审批带着运行类别时照它说、不另取,见 challenge-summary-given
* - ?syn=challenge-approval-unreviewed:正式运行,四句里有两句未能反查(unreviewed:被削弱、经受住的那两句换成数据源
* 已停用、记为无法判断,计数跟着改成无法判断 3 句;工件由检查脚本按同样的改法合成),被推翻的会降档(degrades)
* - ?syn=challenge-approval-unreviewed-over:服务端给的 unreviewed(5)比「无法判断」的句数(1)还多(不该出现):
* 卡上按「无法判断」的句数说,两个数不打架
* - ?syn=challenge-approval-offpath:正式运行,按降档处置,但反方审稿不在带出具契约的成果节点上游(degrades 为假)
* - ?syn=challenge-summary-given:反方审稿那一块单独画,审批自带的运行类别(正式运行)由调用方传进来;检查脚本让这次运行的
* GET 失败,卡上照样写「出具会降档」,也不另发这次请求。审批卡(RunPanel 的 ApprovalCard)把 approval.run_class
* 传给它之后,卡上走的就是这条路
* - ?syn=challenge-approval-dup:同 many,工件里补出来的一句和条目里的一句原文相同、unit 不同(工件由检查脚本合成)
* - ?syn=challenge-approval-norecord:同 many,审稿记录没存上(artifact 为 null)
*
* 运行类别(GET /api/runs/{id})、工件(GET /api/artifacts/{id})都由检查脚本用 page.route 伪造;直接打开时取不到,
* 卡上照实不说降档、点开一句说无法获取。
*
* 检查器:选中反方审稿节点,旁边是同一张图的节点卡片和节点配置的实时内容。
* - ?syn=challenge-inspector:两个上游报告、一个下游报告
* - ?syn=challenge-inspector-chain:「初稿 → 终稿 → 反方审稿」,留空时审直接连着的终稿,反方模型和终稿的模型相同
*/
const report = fx.report as unknown as ApprovalReport
const challenge = fx.challenge as unknown as ApprovalChallenge
/** runClass 为 null:老服务端的审批,不带运行类别 */
function approvalOf(id: string, runId: string, runClass: 'formal' | 'exploratory' | null, reviewed: ApprovalReport): Approval {
return {
id, run_id: runId, node_id: 'review', mode: 'approve', title: '本月月报可以发出吗?', status: 'pending', response: {},
payload: { kind: 'human_node', message: '请审核本月月报。', reports: [reviewed] },
created_at: new Date(Date.now() - 300_000).toISOString().replace('Z', ''),
workflow_name: '月报', node_label: '主管审批', ...(runClass ? { run_class: runClass } : {}),
}
}
function Card({ approval, name, dense = false }: { approval: Approval; name: string; dense?: boolean }) {
return (
<div className={dense ? 'w-[360px] max-w-full border-r bg-panel' : 'mx-auto max-w-3xl p-4'} data-fixture={name}>
<div className="overflow-hidden rounded-lg border" style={{ borderColor: 'var(--warn)' }}>
<ApprovalCard approval={approval} onResolved={() => undefined} />
</div>
</div>
)
}
const RUNS = fx.runs
/** 期 4 起服务端另带的两项:unreviewed(没能反查的句数)、degrades(被推翻的会不会让出具降档) */
type Challenge = ApprovalChallenge & { unreviewed?: number; degrades?: boolean }
const withChallenge = (c: Partial<Challenge>): ApprovalReport => ({ ...report, challenge: { ...challenge, ...c } as ApprovalChallenge })
const many = fx.many as unknown as ApprovalChallenge
/** 没有被推翻的结论:同第一个用例去掉被推翻的那句,计数跟着改 */
const clear: Partial<ApprovalChallenge> = {
counts: { stands: 1, weakened: 1, refuted: 0, inconclusive: 1 },
items: challenge.items.filter((it) => it.verdict !== 'refuted'),
}
/**
* 有两句未能反查:被削弱(u5)、经受住(u8)的那两句换成数据源已停用、没查成,记为无法判断(服务端未能反查的每一句都
* 记为无法判断,unreviewed 不会多于 inconclusive)。原来无法判断的 u11 是查过了、数据不够,不算未能反查
*/
const UNREVIEWED_UNITS = ['u5', 'u8']
const unreviewed: Partial<Challenge> = {
counts: { stands: 0, weakened: 0, refuted: 1, inconclusive: 3 },
items: challenge.items.map((it) => (UNREVIEWED_UNITS.includes(it.unit) ? { ...it, verdict: 'inconclusive' } : it)),
artifact: 'f'.repeat(64),
unreviewed: 2,
degrades: true,
}
// ---------------------------------------------------------------- 检查器
const PROVIDERS = [
{
id: 'pv-main', name: '主力接入', kind: 'anthropic', base_url: null, default_model: 'writer-large',
models: [{ id: 'writer-large' }, { id: 'critic-large' }], enabled: true, extra: {}, api_key_masked: '', has_key: true,
},
{
id: 'pv-spare', name: '备用接入', kind: 'openai', base_url: null, default_model: 'spare-mini',
models: [{ id: 'spare-mini' }], enabled: false, extra: {}, api_key_masked: '', has_key: true,
},
] as Provider[]
const at = (x: number, y: number) => ({ x, y })
const WORKFLOW = {
id: 'wf-challenge-preview', name: '月报(预览)', description: '', tags: [], version: 1, is_template: false, status: 'draft',
graph: {
nodes: [
{ id: 'input', type: 'input', position: at(0, 120), data: { label: '输入', config: { fields: [{ name: 'month', required: true }] } } },
{ id: 'write', type: 'report', position: at(260, 0), data: { label: '写月报', config: { instructions: '写本月月报', model: 'writer-large' } } },
{ id: 'write2', type: 'report', position: at(260, 240), data: { label: '写周报', config: { instructions: '写本周周报' } } },
{ id: 'critic', type: 'challenge', position: at(540, 120), data: { label: '反方审稿', config: {} } },
{ id: 'later', type: 'report', position: at(820, 120), data: { label: '写摘要', config: { instructions: '写一段摘要' } } },
{ id: 'out', type: 'output', position: at(1100, 120), data: { label: '输出', config: {} } },
],
edges: [
{ source: 'input', target: 'write' }, { source: 'input', target: 'write2' },
{ source: 'write', target: 'critic' }, { source: 'write2', target: 'critic' },
{ source: 'critic', target: 'later' }, { source: 'later', target: 'out' },
],
},
} as unknown as Workflow
/** 链式接法:初稿、终稿都在上游,直接连着的只有终稿。终稿和反方审稿写的是同一个模型 */
const CHAIN = {
id: 'wf-challenge-chain', name: '月报(链式,预览)', description: '', tags: [], version: 1, is_template: false, status: 'draft',
graph: {
nodes: [
{ id: 'input', type: 'input', position: at(0, 120), data: { label: '输入', config: { fields: [{ name: 'month', required: true }] } } },
{ id: 'draft', type: 'report', position: at(260, 120), data: { label: '初稿', config: { instructions: '写初稿', model: 'spare-mini' } } },
{ id: 'final', type: 'report', position: at(520, 120), data: { label: '终稿', config: { instructions: '改成终稿', model: 'writer-large' } } },
{ id: 'critic', type: 'challenge', position: at(780, 120), data: { label: '反方审稿', config: { model: 'writer-large' } } },
{ id: 'out', type: 'output', position: at(1040, 120), data: { label: '输出', config: {} } },
],
edges: [
{ source: 'input', target: 'draft' }, { source: 'draft', target: 'final' },
{ source: 'final', target: 'critic' }, { source: 'critic', target: 'out' },
],
},
} as unknown as Workflow
const nodeTypes = { card: NodeCard }
/** 同一张图的节点卡片:看反方审稿节点的摘要跟着配置变 */
function MiniCanvas() {
const nodes = useStudio((s) => s.nodes)
const edges = useStudio((s) => s.edges)
return (
<ReactFlowProvider>
<ReactFlow nodes={nodes} edges={edges} nodeTypes={nodeTypes} fitView nodesDraggable={false} proOptions={{ hideAttribution: true }} />
</ReactFlowProvider>
)
}
function ConfigView() {
const config = useStudio((s) => s.nodes.find((n) => n.id === 'critic')?.data.config)
return (
<pre className="mono max-h-40 overflow-auto whitespace-pre-wrap border-t p-2 text-2xs text-dim [overflow-wrap:anywhere]"
data-challenge-config="">
{JSON.stringify(config ?? null, null, 1)}
</pre>
)
}
function InspectorCase({ name, workflow }: { name: string; workflow: Workflow }) {
// 首次渲染前装好图和接入:检查器、卡片一挂上就是选中反方审稿节点的样子
useState(() => {
useCatalog.setState({ providers: PROVIDERS })
useStudio.getState().load(workflow)
useStudio.getState().select('critic')
return true
})
return (
<div className="flex min-h-screen flex-col md:h-screen md:flex-row" data-fixture={name}>
<div className="w-full shrink-0 border-b bg-panel md:h-full md:w-[360px] md:border-b-0 md:border-r" data-challenge-inspector="">
<Inspector />
</div>
<div className="flex min-h-0 min-w-0 flex-1 flex-col">
<div className="h-72 min-h-0 md:h-auto md:flex-1"><MiniCanvas /></div>
<ConfigView />
</div>
</div>
)
}
export const cases: Record<string, () => JSX.Element> = {
'challenge-approval': () => (
<Card name="challenge-approval" approval={approvalOf('ap-chal-1', RUNS.formal, 'formal', withChallenge({}))} />
),
'challenge-approval-flag': () => (
<Card name="challenge-approval-flag"
approval={approvalOf('ap-chal-2', RUNS.flag, 'formal', withChallenge({ on_refuted: 'flag' }))} />
),
'challenge-approval-explore': () => (
<Card name="challenge-approval-explore"
approval={approvalOf('ap-chal-3', RUNS.explore, 'exploratory', withChallenge({ on_refuted: 'degrade' }))} />
),
'challenge-approval-none': () => (
<Card name="challenge-approval-none" approval={approvalOf('ap-chal-4', RUNS.none, 'formal', report)} />
),
'challenge-approval-many': () => (
<Card name="challenge-approval-many" approval={approvalOf('ap-chal-5', RUNS.many, 'formal', { ...report, challenge: many })} />
),
'challenge-approval-broken': () => (
<Card name="challenge-approval-broken"
approval={approvalOf('ap-chal-6', RUNS.broken, 'formal', withChallenge({ artifact: 'e'.repeat(64) }))} />
),
'challenge-approval-dense': () => (
<Card name="challenge-approval-dense" dense approval={approvalOf('ap-chal-7', RUNS.formal, 'formal', withChallenge({}))} />
),
'challenge-approval-clear': () => (
<Card name="challenge-approval-clear" approval={approvalOf('ap-chal-8', RUNS.clear, 'formal', withChallenge(clear))} />
),
'challenge-approval-clear-flag': () => (
<Card name="challenge-approval-clear-flag"
approval={approvalOf('ap-chal-9', RUNS.clearFlag, 'formal', withChallenge({ ...clear, on_refuted: 'flag' }))} />
),
'challenge-approval-explore-flag': () => (
<Card name="challenge-approval-explore-flag"
approval={approvalOf('ap-chal-10', RUNS.exploreFlag, 'exploratory', withChallenge({ on_refuted: 'flag' }))} />
),
'challenge-approval-runfail': () => (
<Card name="challenge-approval-runfail" approval={approvalOf('ap-chal-11', RUNS.runfail, null, withChallenge({}))} />
),
'challenge-approval-dup': () => (
<Card name="challenge-approval-dup"
approval={approvalOf('ap-chal-12', RUNS.dup, 'formal', { ...report, challenge: { ...many, artifact: 'd'.repeat(64) } })} />
),
'challenge-approval-norecord': () => (
<Card name="challenge-approval-norecord"
approval={approvalOf('ap-chal-13', RUNS.norecord, 'formal', { ...report, challenge: { ...many, artifact: null } })} />
),
'challenge-approval-unreviewed': () => (
<Card name="challenge-approval-unreviewed"
approval={approvalOf('ap-chal-14', RUNS.unreviewed, 'formal', withChallenge(unreviewed))} />
),
'challenge-approval-unreviewed-over': () => (
<Card name="challenge-approval-unreviewed-over"
approval={approvalOf('ap-chal-16', RUNS.over, 'formal', withChallenge({ unreviewed: 5, degrades: true }))} />
),
'challenge-summary-given': () => (
<div className="mx-auto max-w-3xl p-4" data-fixture="challenge-summary-given">
<ChallengeSummary reports={[withChallenge({ degrades: true })]} runId={RUNS.given} runClass="formal" />
</div>
),
'challenge-approval-offpath': () => (
<Card name="challenge-approval-offpath"
approval={approvalOf('ap-chal-15', RUNS.offpath, 'formal', withChallenge({ on_refuted: 'degrade', degrades: false, unreviewed: 0 }))} />
),
'challenge-inspector': () => <InspectorCase name="challenge-inspector" workflow={WORKFLOW} />,
'challenge-inspector-chain': () => <InspectorCase name="challenge-inspector-chain" workflow={CHAIN} />,
}