[loopx-agent]

Goal And Delivered Outcome

新运行来源的研究历史首次读取失败时，页面一条记录都没有，却仍提示“已读记录保留”。现在中英文提示明确指出当前来源无法读取；同源刷新失败仍保留有效记录及原有保留提示。

把已公开的真实 R7 context on/off 历史接入现有前端 smoke 集合，覆盖两源共享同一逻辑 Goal ID 时的失败、重试、原文下载和返回。来源隔离在基线上已经通过；这次修复的是首次失败提示，并补上真实数据的跨源恢复验证。

Author Declaration

• Written by: model_agent — OpenAI Codex.
• Specification: docs/development/frontend-delivery.md#task-first-delivery--从用户任务开始 and docs/development/design.md#populated-page-review, revision 33b15d037191be7273e39e1e78a01e66b316d732; the accepted request is to use real benchmark display state as a core frontend reference.

Criterion	Disposition	Implementation / decisive check
Distinguish unavailable, unknown and retained data	implemented	Existing GoalResearchHistory zero-record error branch; failing-before/passing-after semantic assertion
Preserve real records and source identity	implemented	Existing public R7 captures, unchanged Goal IDs, exact exported IDs/order/text/time/status/measurements, unchanged fixture bytes
Packaged user journey and recovery	implemented	Actual RuntimeWorkspaceServer HTTP/projection owner; source switch, keyboard retry, Back/Forward, bilingual 1440/390 views, zero non-GET requests

Scope And Continuation

Complete within this bounded copy/recovery slice. Reuses the existing Explore/Research owner and fleet browser; no new provider, configuration, state vocabulary or effect authority. The Python fixture adapter supplies captured records and controlled availability to the actual HTTP owner; it does not qualify live SSH/CLI source access, current solver outcomes or an installed App. Broader research work remains separate.

Bounded refactoring pass: reused the existing scenario catalog/browser and preserved the reader implementation, whose source isolation already passed. No model, solver, evaluator, benchmark launch, installation or release was performed.

Validation

• Tested revision: 7e5530d2b6c2670afa15f3394d21938ebdf7f530 (base 33b15d037191be7273e39e1e78a01e66b316d732, target main).
• Run state: finished.
• Input classes: public_fixture, synthetic (catalog shell and controlled fault).

Check kind	Result	Evidence / limitation
regression_parity	passed	New zero-record preservation assertion fails on baseline 4b7afef; passes after the copy fix. Existing source reset behavior passes on both.
static	passed	Verified packaged Chat build, Dashboard typecheck and strict TypeScript check of the new scenario.
integration	passed	25 tests in test_runtime_workspace.py, test_goal_source.py, test_explore_history.py.
real_entrypoint	passed	Packaged research-source-recovery fleet scenario: on 33 records → same-source failed refresh retains 33 → off first-read failure exposes no old graph → keyboard retry reads 11 → original JSON export → Back/Forward. Both languages; narrow overflow checked.
regression_parity	passed	Existing runtime-source-workspace-browser-smoke.mjs synthetic lifecycle/analysis-failure checks.
manual	passed	Ego inspection of default real history, expanded scoped negative result, first-read error and keyboard recovery at desktop/390 widths.
static	failed	Optional maintainability ratchet: unchanged oversized _admitted_interaction_next_cli_actions in interaction_contract.py. Immutable base/head runs with the same command/environment reproduce identical complete failure output. This remains a failed check.
static	passed	Standard premerge gate: four direct checks, three passing catalog checks and public-boundary scan; the ratchet above is reported as an inherited advisory. Author CQR records the exact scope and baseline comparison; it is not independent PR approval.

Reproduce after building Chat:

LOOPX_PERSONAL_WORKSPACE_PACKAGED=1 LOOPX_PERSONAL_WORKSPACE_SCENARIO=research-source-recovery LOOPX_PLAYWRIGHT_PACKAGE="$PWD/apps/presentation/dashboard/node_modules/playwright" node --experimental-strip-types examples/personal-workspace-browser-smoke.mjs

CI and independent review remain separate from the local evidence. No merge authorization is claimed.

Frontend / Visual Evidence

Packaged UI with original public R7 records and a disposable read-only adapter. Before: 4b7afef; after: tested head above. No live/private session screenshots. Captures remain outside Git.

Before / 首次失败错误声称保留记录（390px）：

After / 明确当前来源未读到；保留原重试入口（390px）：

Default populated first viewport / 原始 on 历史 33 条（1200px）：

Expanded original negative result / 406 → 406 的有限范围负面证据（1200px）：

Keyboard recovery / off 历史恢复为 11 条（390px）：

Attention/craft review: the source name, unread coverage and one retry control answer the failure in place. The misleading preservation claim disappears only when no records exist. Default history and a long negative finding remain readable; no layout, color, graph semantics or source-authored outcome label changed. This is a packaged fixture qualification, not a claim of benchmark benefit or installed behavior.

Type of Change

• Bug fix
• Test update

LoopX Area

• Public docs or presentation surface

[loopx-agent]

Reviewer: model_agent; gpt-6.1-sol; OpenAI; runtime_reported; reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机 / Motivation

在个人工作区阅读研究历程、切换运行来源的用户。 切到一个尚未成功读取的来源并遇到网络错误时，旧版声称“已读记录保留”；新版明确指出此来源读取失败，用户可以直接重试。 首次失败不会暗示存在已保留的记录；同来源刷新失败仍保留已有记录，重试后能继续阅读正确来源。 本 PR 不证明 benchmark 改善、模型效果、SSH 实时可用性或已安装桌面 App 的交付。

问题在错误提示的真实性，而非缺少另一个读取系统。保留旧提示会让用户把没有读到的数据误认为仍可阅读，增加切换来源后排查数据归属的成本。这里两条文案就是完整、可回滚的修复；验证不能仅检查字符串，还要真正制造读取失败。

改动思路 / Approach

现有 GoalResearchHistory 继续拥有显示选择：只有 read.events.length 非零时，才显示“刷新失败，当前记录已保留”；零记录的失败走新的来源提示。来源 URL、分页、缓存失效、重试、下载和后台读取仍复用现有 owner。无新增模式、持久状态、权限或配置。

生产修复留在现有投影分支，真实 HTTP 与记录归属仍由原读取 owner 负责。当前 PR 的交付边界是未读来源的准确错误文案，以及围绕既有恢复链路的一条可持续回归场景。增加新的错误类型或重写缓存会扩大成本，不能改善这次已明确的错误。

实际触发	本次检查的结果
已有 33 条记录后刷新失败	保留 33 / 33，继续提示已保留；键盘重试恢复
切换到同 Goal 的另一物理来源但首次读取失败	显示此来源失败，不再显示旧来源的 33 条记录
恢复第二来源	11 / 11，下载为该来源的 11 条原记录

具体改动 / Changes

以接受本次改动前的 docs/development/frontend-delivery.md#task-first-delivery--从用户任务开始 和 docs/development/design.md#populated-page-review 为规格，spec_revision 为 33b15d0。映射：task-first-delivery 已实现，读取失败后可在同页恢复并读回来源；populated-page-review 已检查，使用公开 R7 on/off 的长标题、负面观察、来源详情与不完整状态。此处是按章节命名的标准映射，不另造产品验收编号。

关键代码讲解

• GoalResearchHistory 的文案：中文和英文各替换一个值，总计 2 加 2 删。新的主句只用于尚无记录的失败，不承诺不存在的快照。
• 错误展示与重试分支：已有记录时仍走专用保留提示；详情与 refresh 按钮保持可达。因此缩短主句没有删除重试操作和错误详情。
• researchSourceRecoveryScenario：152 行场景复用现有公开捕获和 RuntimeWorkspaceServer，只控制传输可用性。注册入口增加 2 行；该场景覆盖保留、跨源、键盘、导出与历史导航，未引入并行生产状态机。

PASS：不可变基线和 head 的后端集成都为 25 项；两边 control-plane typecheck、Chat bundle build 均通过。新增场景用仓库 Node 类型完成 strict TypeScript 检查。最初未带 Node 类型根的临时编译命令失败，改正编译环境后通过，保留该观察。

我用 Ego 操作各自构建的基线和 head，通过实际 HTTP owner 独立重现中英文首次失败、同来源保留、切换后的重试、11 条记录下载及 Back/Forward。导出保留 ID、顺序、正文、数值和覆盖状态；缺失字符串按既有投影归一为空串。最后一次实际失败/重试记录为 GET 503 后 GET 200，没有写请求。公开输入文件与对应 Git blob 相同。

对主干的风险 / Risks

未发现阻塞问题。最强反例是切换两个同 Goal、不同物理来源时把旧记录保留成新来源的完整历史；独立页面检查显示新来源标题、来源未读取状态和空历史，恢复后读到其 11 条记录。另一个方向也检查了：保留 33 条有效记录时没有被新的主句误报为全部不可读。

PASS：标准 source-checkout premerge 的必需检查及公共边界检查通过。FAIL：可维护性 ratchet 在不可变基线和 head 均以同一完整诊断失败，指向未修改的 _admitted_interaction_next_cli_actions。它是原有 advisory，不改成 PASS；独立比较了完整失败输出、输入集合和运行环境，保留该债务。

本次没有 opt-in/default-off 声明，常规读写、调度、配额与已安装指引未变。主要限制是使用既有公开捕获控制传输错误，未连接真实 SSH 或生产运行来源；仓库自带的 Playwright 启动场景没有另行执行，独立浏览器验证由 Ego 完成。首次视口证据在 PR 描述中；我另看了实际打包页面的桌面与窄屏状态。

我的整体评价 / Verdict

对这个文案缺陷，交付判断为 goal_achieved；更大的 benchmark 与安装验收仍在本 PR 边界之外。long_horizon：preserved，重试、导航、原记录下载可继续工作，没有新增循环或承诺。user_experience：improved，首次失败的描述真实且重试可达。

实际页面的 1200×888 默认/展开视图和 390×844 错误/展开视图已检查：来源、覆盖与重试有清晰层级；长标题在窄屏换行，正文与主数值可读，完整原文仍可展开。已展开的正文和来源字段可继续阅读；未验证范围是已安装 App、SSH 与 live benchmark，不能从捕获记录外推。自动 overflow 检查仅作补充，不替代视觉判断。

该修复沿用现有 vocabulary 和 owner，生产增量适中，无必要的兼容分支清理。没有把全部父目标完成作为批准条件。APPROVE 是这个精确 head 的代码结论；独立 merge readiness 和维护者合并授权仍须满足。

English verdict: APPROVE 7e5530d — no blocking finding; first-read copy and recovery verified with packaged UI, real HTTP owner and local checks. Existing advisory debt and installed/live-source limits remain.