Repository navigation
test(workspace): preserve remote source authority readback through current help - #7362
Conversation
Signed-off-by: BigDataDZ <76271875+BigDataDZ@users.noreply.github.com>
loopx-agent
left a comment
There was a problem hiding this comment.
Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | reasoning_effort=max
动机
运行远端来源浏览器验收的贡献者。原场景等待已退役的标签而超时,后续只读权限与移动端断言无法执行;修正后从现有帮助入口读取完整边界,再继续原场景。本次独立重现基线旧标签超时,修正版本完整打包场景通过。不修改产品界面或 SSH 权限,不证明真实 SSH、安装态采用或整个浏览器分片通过。
改动思路
沿用已有 readContextHelp,从 .personal-read-only-notice 中以精确可访问名称打开“只读来源”,核对完整读写边界,再执行既有 Escape、隐藏状态和焦点返回检查。删除权限断言或重新加回产品说明都会扩大或削弱边界,当前做法只修复已有测试入口。产品默认行为、来源配置和写入权限没有变化。
具体改动
关键内容讲解
完整 diff 仅 navigation-sorting.mjs +2/-1。navigationSortingScenario 把退役标签等待替换为现有 helper 调用;readContextHelp 继续验证“可查看 Goal、Task、证据与运行状态,写入、纠偏和 Agent 会话留在来源主机”的完整说明及键盘恢复。后续 SSH 来源切换、隐藏写控件、Task/调度、设置返回、移动端和真实持久排序断言均保留。
规范:good-smokes.md,spec_revision e905d3433ce01c6cecfe96ce9655d3c095aa1c4c。Start With The Contract、Review Checklist、Validate The Change 三项均 implemented;它们是规范标题定位符。另已核对同版本 design.md 的 Explanations On Demand,产品原本就把完整说明放在上下文帮助中。
对主干的风险
无阻塞发现。独立执行同一打包场景:基线在旧精确标签处超时,当前 head a0af34fc975a5fdb4a8d2468851957d7329c5629 完整通过,未放宽超时或删断言。当前标准 premerge 的 4 项直接检查及 3 项选中检查通过。首次 premerge 因本地工作树缺少根目录 TypeScript 依赖失败;只恢复忽略依赖后同 head 整体重跑通过,保留原失败。没有查询或等待 CI。
核对了同作者近期八项变更,远端权限、研究报告导航和研究 HTTP fixture 各有独立不变量,本次复用原场景,不构成重复 smoke。打包界面和浏览器为真实执行,远端 API 内容为合成数据。
我的整体评价
APPROVE。命名切片已完成;long_horizon=preserved,运行时与持久状态规则不变。未改动产品渲染,已检查远端只读提示与 ContextHelp 的实际调用;无需新增视觉验收。真实 SSH 与安装态采用未验证。原有 helper 已是最小修复,无需新增抽象。评审通过不授予合并权限,也不代表完整浏览器分片或原生 Windows 被本次独立验证。
English verdict: APPROVE at a0af34f. The existing semantic help and focus checks replace a retired label without weakening downstream authority coverage. The full packaged scenario and current premerge pass; the immutable baseline reproduces the old timeout. Native Windows, live SSH and installed adoption are not independently qualified here.
Problem and result
The required browser lane now times out at a retired remote-projection label in navigation-sorting. The shipped remote-source notice exposes its complete reading/writing boundary through contextual help.
Read the existing 只读来源 help instead. Verify that Goals, Tasks, evidence and run status are readable while writes, guidance and Agent sessions remain on the source host. Reuse the scenario's existing Escape/focus helper and preserve all downstream SSH source, hidden write-control, task/schedule and mobile checks.
Validation
Written by model_agent (OpenAI Codex), under the author's direction. Basis: current-main browser failure and docs/development/design.md explanations-on-demand. Target/merge base: main e905d34. Tested revision: a0af34f (full SHA in remote head). Other current browser failures remain separately owned; this does not claim a whole-shard pass. Current-head CI and independent review pending. No Goal mutation, model, benchmark, global upgrade or self-merge.