Repository navigation
fix(explore): retain execution feedback in product readers - #9071
Conversation
Signed-off-by: lusendong.6789 <[email protected]> Co-authored-by: TRAE CLI <[email protected]>
Signed-off-by: lusendong.6789 <[email protected]> Co-authored-by: TRAE CLI <[email protected]>
fa29e17 to
2d889f9
Compare
cocolord
left a comment
There was a problem hiding this comment.
Reviewer: model_agent; self_reported; model=GPT-6; provider=OpenAI; runtime observation unavailable (session_not_bound).
Approval conclusion (author-owned PR; GitHub blocks formal self-approval)
动机 / Motivation
阅读已保存 Explore 结论并检查方法与结果关联的 App/Lark 用户,会遇到同一来源已返回执行观察、保存结论页却丢弃详情的问题。同一来源返回已记录执行观察时,旧保存结论页丢弃详情,新页保留原方法、实际方法、改向原因、成本及结果引用。App 保存结论、研究历史及既有 Lark 摘要保留相同源事实,缺失测量仍是未知。不启动方法、激活能力、确认模型采用、证明收益或关闭整个机制 RFC。安装态升级/读回、剩余执行切换和独立成果/完整成本验收继续开放。
改动思路 / Approach
既有验证 owner 和一个按需 reader 足以修复遗漏;Lark 复用原行投影,无新增状态、执行、权限或自动采用。当前 PR 完成保存结论/历史/Lark 的来源读回与打包资格;安装态与独立采用/收益另沿原验收。保持现状会继续丢失保存结论的执行来源,另建 API 或 renderer 会重复既有语义。当前实现直接复用 Explore codec,并把历史展示抽成一个组件;原日志、结果有效性、历史顺序、认领和结算仍由原 owner 负责。
具体改动 / Changes
规范基础为 docs/architecture/rfcs/mechanism-convergence-and-learning-loop-v0.md,revision 016e198230b87b1991c73efca48b622df319f6dd;判断使用变更前已接受的条款,而非此 PR 新增 checkpoint。Product readback 在保存结论、历史和 Lark 来源投影中实现;Stale, conflicting, late and unknown evidence 通过错误候选拒绝及 unknown 保留验证;Off, failure, older Hosts and recovery 保留省略字段、关闭执行后的读回和读取恢复;Decisions and real execution 中真实模型采用及独立结果仍为 deferred。
关键代码讲解
exploreResultPageSchema / fetchExploreResults保留可选观察,并复用normalizeExploreExecutionObservation。缺失仍合法;错误候选进入原读取错误路径,没有新状态分类器。ExecutionObservationReadback被保存结论和历史共用。默认关闭,null 显示“未测量”,输入版本与候选/结果引用保留;来源记录不升级成方法有效或成果验收。_finding_summary / _execution_summary只扩展原 Summary。无观察的结论逐字保持原摘要,有观察时追加同源方法、成本和引用;Status、row lineage 及外部写入权限未变。
fixture 和 smoke 使用 13 条研究记录、12 条结论与重复长文本;真实 File/SQLite 后端提供 HTTP 和 packaged UI。新增四张 PNG 是当前指南用于入口、详情、窄屏引用及旧版遗漏排查的规范图示,PR 复用同一份图,运行截图仍忽略。内容、fixture 来源、IHDR/IDAT/IEND 元数据和远端字节均已检查。
对主干的风险 / Risks
最强反例是先加载合法结论,再返回另一个候选的结果,仍留下旧执行事实。当前 packaged 负例只改变传输,codec 拒绝后原 reader 清空旧页、显示错误,修复来源后刷新恢复;两种真实后端的 source_unchanged 和零 mutation assertions 均通过。11 个 HTTP/Lark 用例、两种 packaged reader、TypeScript/Chat build 及标准 premerge 的 6 个直接、19 个执行检查通过。premerge 原状态仍为 manual_review_required;四张二进制图片的内容、元数据、来源在本次 review 人工核验,不将自动文本扫描冒称图片隐私证明。
关闭 Explore 时,已保存证据依旧可读;未带观察的保存结论和 Lark 摘要保持原表现,没有新增配置或确认轮次。远端文档构建失败来自未改动的 turn-envelope benchmark 链接;merge-gate 同一运行的 changes 任务被取消,仍有 pending/cancelled 检查。这些状态保留为独立 merge-readiness 门槛,未声明远端 CI 全绿。当前验证不覆盖已安装客户端或实际 Lark 外部写入;合成观察不证明来源测量真值、模型采用或方法收益。
我的整体评价 / Verdict
对长期工作,保留原记录、认领/扣额和继续执行权限,仅恢复原来源的可读性;对用户旅程,计划/实际方法、改向及成本现在在已保存结论内可找到,默认折叠保留结论层级。13 条研究记录、12 条长/重复结论及 refuted/unknown 混合状态已检查;默认结论和展开执行细节分别可读,长引用在窄屏内换行。已安装客户端及实际 Lark 外部写入未验收;来源观察不证明模型采用或方法收益。
全范围判断是 proportionate 的 justified increment:复用原语义和投影,不新增共享 vocabulary,观察字段仍可省略。当前 PR 完成保存结论/历史/Lark 的来源读回与打包资格;安装态与独立采用/收益另沿原验收。精确评审头 2d889f95b701dae8d7bb70612f64e28c39202c40 无已验证的 blocking finding;这里发布 author-owned COMMENTED approval conclusion,既不关闭整个 RFC,也不授予 merge 权限。
English verdict: APPROVE for this bounded readback slice, published as an author-owned COMMENTED review. Current local evidence covers the changed invariants; CI/merge readiness, installed adoption and independent outcome/full-cost qualification remain separate open gates.
|
This pull request has merge conflicts with Choose the remote for the base repository, not an out-of-date fork. git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEADFor a same-repository clone whose Keep the DCO |
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent · gpt-6.2-sol · OpenAI · runtime_reported · xhigh
动机 / Motivation
重新打开 Explore 已保存结论的用户需要复核实际方法和成本。 后端原记录有执行观察,旧 App 读回会丢掉它,Lark 也不显示。 当前打包页面的 12 条结论中,可展开看到计划与实际方法、12.5/60 秒,以及两项未测量值。桌面和 390px 窄屏均可查看,页面没有横向溢出。 本次交付只读回原执行事实,实际模型采用、方法收益、客户端安装态仍待验收。
改动思路 / Approach
App 两个阅读入口共享展示组件,Lark 仍投影同一经过校验的记录。 这次只改变现有结论与历史界面的读回和 Lark 摘要。 该组件和用户说明保留了记录事实、负结果和独立验收之间的界限。 两条路径均复用原有来源和相同身份。
具体改动 / Changes
审查提交 2d889f9,基线 016e198。17 个文件,+633/-20。机制 RFC 的反馈读回、未知与拒绝由原生入口实测核对。
关键代码讲解
新增 ExecutionObservationReadback 供已保存结论与历史共用。Lark 的 _execution_summary 读取同一观察,并保留未知值。 App 读回 Lark 投影
新路径复用原有来源和类型判断,测试通过常规入口观察完整结果。生产代码没有另建来源、写回或结果状态。旧记录继续由原接口读取,验收仍依据原始任务和当前头的证据。成对 RFC 记录此处可验证的边界与后续采用范围。
对主干的风险 / Risks
无阻断发现。File 和 SQLite 的 HTTP/Lark 验证拒绝错误候选引用,禁用 Explore 后仍能读原结论。无执行观察的记录保持原样。 11 项 File/SQLite HTTP 与 Lark 测试通过。两种打包浏览器场景通过。Ego Browser 检查实际本地阅读页,独立核对桌面和 390px 的默认与展开截图。 快速预合入的直接和选定检查均无失败。四张指南 PNG 均与准确头字节一致,块校验通过,保留工具的四项截图人工门禁。 保留当前头的测试事实和原有拒绝条件。远端 CI 未查询。当前平台合入存在冲突,仍须独立解决。
我的整体评价 / Verdict
批准本次切片。原始结果和权威保持可追溯。实际运行时采用及后续成果仍需用现有工作流核查。默认视图以结论正文为主,细节按需展开。窄屏把指标排成纵列,未测量和值分开,桌面哈希可折行。 已安装客户端的更新及读回、真实模型消费和成果全成本未测。
English verdict: APPROVE at 2d889f9. The existing execution facts reach the packaged readers without creating measurements or acceptance.
Signed-off-by: LoopX Agent <[email protected]>
Saved Explore findings already return execution observations, but the App decoder discards them and the Lark summary omits them. This change retains the existing typed observation, shares one disclosure between findings and history, and presents the same execution facts in the existing Lark summary. Unknown measurements remain unknown; absent observations retain their earlier display.
保存的 Explore 结论已由后端返回执行观察,但 App 解析会丢弃它,Lark 摘要也缺少这些信息。本变更保留原类型校验,结论和历史共用一个按需展开的读回,Lark 同步展示计划/实际方法、改向原因、成本及结果引用。记录本身不证明模型采用、方法收益或成果验收。
The illustrated Explore guide explains the App entry, reading measured versus unknown values, narrow-screen references, and diagnosis of the earlier saved-reader omission. Its four public-safe images are canonical guide illustrations, reused below; transient capture output remains ignored. The bilingual mechanism RFC retains this bounded source checkpoint and its remaining acceptance gaps.
Validation
2d889f95b701dae8d7bb70612f64e28c39202c40; verified targetmain:016e198230b87b1991c73efca48b622df319f6dd.1a3ab75a53107b29ee1b8bff56e2b8263259dec6) bundle reproduces the saved-finding omission on both providers with the same backend and synthetic inputs; the candidate restores execution detail.manual_review_requiredfor four binary guide images: exact-head review must inspect their content, metadata and source. All images were visually inspected; PNGs contain only IHDR/IDAT/IEND chunks, and guide references resolve. DCO and bilingual gates passed. This is not merge readiness or authority.Visual evidence / 视觉证据
Packaged UI, synthetic populated data (13 research events, 12 findings), isolated SQLite backend; the same scenarios also passed on File. These images were visually inspected and contain no private runtime state. The default research screen retains on-demand detail; the saved finding adds the omitted execution disclosure.
Default desktop reading entry (1280×900):
Before: the target-base saved-finding reader omits the observation returned by the same source:
After: expanded saved-finding detail from the candidate, with measured time and unknown iteration/stop values:
Narrow detail at 390px wide, captured in a taller 1400px reading viewport to include the complete block. The actual 390×844 default/expanded viewport and no-horizontal-overflow assertions also passed; its one-run captures remain ignored:
Long, repeated findings and refuted/unknown states were inspected. The complete visible saved-reader composition was inspected separately from the initial viewport. Guide readers use these images to locate the detail, distinguish missing measurements, read wrapped references, and diagnose the older omission. They are not a capture archive. Tested source is the exact head above; the before image uses the pinned comparison-base bundle. Canonical guide images were captured before the final rebase with byte-identical product reader/fixture sources; the latest-head packaged replay passed, and both detail images remain byte-identical to that replay. Installed-client upgrade/readback remains untested.
复现:在 dashboard 目录运行
npm run build:chat,再从仓库根运行node apps/presentation/dashboard/smoke/execution-feedback-browser-smoke.mjs。截图区分默认首屏、展开详情和完整窄屏读回;详情图不冒充默认首屏。指南保留单份规范图片,运行截图仍忽略。真实打包界面及隔离后端已验证,已安装客户端仍未验收。Remaining mechanism work
The whole RFC remains partial: execution convergence and remaining caller migration, actual adoption, calibration, cross-task reuse, and independent outcome/full-cost qualification remain open. EdgeBench is later evaluation and does not block independent engineering. This proposal grants no execution, activation or merge authority.