Files
proxycast/docs/test/harness-evals.manifest.json
T
2026-03-27 14:35:56 +08:00

90 lines
3.2 KiB
JSON

{
"manifestVersion": "v1",
"title": "Lime Harness Eval Manifest",
"defaults": {
"requiredArtifacts": [
"input.json",
"expected.json",
"grader.md",
"evidence-links.json"
],
"requiredInputFields": [
"session.sessionId",
"session.threadId",
"task.goalSummary",
"classification.suiteTags",
"classification.failureModes",
"linkedArtifacts.handoffBundle.relativeRoot",
"linkedArtifacts.evidencePack.relativeRoot"
],
"requiredExpectedFields": [
"successCriteria",
"blockingChecks",
"artifactChecks",
"graderSuggestion.preferredMode"
],
"requiredEvidenceFields": [
"handoffBundle.relativeRoot",
"evidencePack.relativeRoot"
]
},
"suites": [
{
"id": "repo-fixtures",
"title": "仓库固定 Replay 样本",
"priority": "P0",
"roadmap": "P3-2 Eval runner",
"description": "固定一条可在 CI 和 nightly 中稳定运行的 replay fixture,先验证 grader 合同与样本结构,而不是等真实工作区样本才能开始。",
"upstream": {
"codex": "沿用 Codex evidence-first 的 replay / grader 形状。",
"aster": "沿用 Aster runtime / thread / turn 的会话事实边界。",
"lime": "把样本、脚本和摘要统一落在 Lime 仓库 current 主线。"
},
"cases": [
{
"id": "fixture-minimal-pending-request",
"title": "最小 pending request Replay 样本",
"source": "repo_fixture",
"caseDir": "docs/test/harness-fixtures/replay/minimal-pending-request",
"tags": ["conversation-runtime", "replay", "handoff", "evidence"]
}
]
},
{
"id": "repo-promoted-replays",
"title": "仓库沉淀 Replay 样本",
"priority": "P1",
"roadmap": "P3-6 Replay 样本沉淀",
"description": "把工作区导出的高价值 replay case 提升为仓库 current 样本,用于固定回归入口与 nightly 趋势对比。",
"upstream": {
"codex": "沿用 Codex 把真实失败沉淀为 replay 资产的做法。",
"aster": "继续复用 Aster runtime 导出的 thread / turn / evidence 边界。",
"lime": "由 Lime 持有 promotion 命令、repo fixture 目录与 current manifest。"
},
"cases": []
},
{
"id": "workspace-replay-discovery",
"title": "工作区 Replay 自动发现",
"priority": "P1",
"roadmap": "P3-2 Eval runner",
"description": "扫描当前工作区 `.lime/harness/sessions/*/replay`,把真实会话导出的 replay case 接进固定摘要格式。",
"upstream": {
"codex": "参考 Codex 把真实失败沉淀为 replay / eval 样本的习惯。",
"aster": "复用 Aster session runtime 导出的 thread / turn 上下文。",
"lime": "让 Lime 的 handoff bundle 与 evidence pack 成为 eval 的事实源。"
},
"cases": [
{
"id": "workspace-session-replays",
"title": "工作区会话 Replay 样本",
"source": "workspace_replay_discovery",
"root": ".lime/harness/sessions",
"allowZeroMatches": true,
"tags": ["workspace", "replay", "runtime-export"]
}
]
}
]
}