当前判断
- 框架主干已经完成到 explicit executorReceipt / terminalResult / executorArtifacts 边界
- 真实 browser executor 不需要重写 terminal backwrite，只需要替换生产者
- D1 继续做 run / attempt / ledger / audit 的 source of truth
- R2 failure evidence 首刀已经落地；真实 screenshot binary 仍留给 real runner 接入时替换
- POST /api/runs/pickup-next 与 browser-runner-pickup.mjs 已把真实 runner intake 的 pickup / failed / success proof 接到同一条主线上
默认推进顺序
- Slice A：pickup / fail / run attempts
- Slice B：失败证据 contract
- Slice C：retry / reservation contract
- Slice D：R2 失败证据落盘
- Slice E：真实 Node + Playwright runner
本次已落地
- workflow_run_attempts 已进入 schema
- POST /api/runs/:runId/pickup / fail / retry / complete 已进入代码
- POST /api/runs/pickup-next 已进入代码，并允许 runner 直接消费 queued run intake
- GET /api/runs 现在会返回 recoveryQueue / recoveryStats，Runs workbench 已独立显示 manual recovery queue
- POST /api/runs/:runId/retry 现在支持 retryMode=manual_recovery，让 owner/operator 可以在保留 reservation 的前提下手动重排
- POST /api/runs/:runId/complete 现在支持 browser-runner success payload，并把 executorReceipt / sinkWrites / latestAttempt success detail 持久化到 accepted run truth
- Runs workbench 已能显示 running / failed、latest attempt、retry summary 与 failure evidence
- EVIDENCE_R2_BUCKET -> toolplane-ai-private 已进入 staging 绑定
- GET /api/runs/:runId/evidence/:evidenceId 已能读取持久化后的 failure evidence
默认下一刀
- 更重的 selector / screenshot 证据模型
- 继续复用现有 pickup / fail / retry / complete 合同，而不是重写 terminal backwrite
- 让 real runner 从当前 proof client 继续推进到独立 executor 进程与更重的 failure evidence