[loopx-agent]

Problem and result

A Goal can load normally while its Codex execution badge stays unavailable: a verified but old observation cache makes the reader replay a large intervening transcript before reaching a recent native turn marker. The cache can also record the current file extent while forward catch-up is incomplete.

The existing read-only provider now searches the latest tail when a loaded checkpoint or large append exceeds the remaining read budget. An incomplete tail still stays unknown and continues its bounded scan. Only caught-up observations replace the durable cache checkpoint. The shared presentation says “执行状态未知 / execution state unknown”, so unavailable execution observation does not imply unreadable Goal content.

Contract and scope

• Basis: the existing bounded local Codex observation contract and observed liveness design rule.
• Entry points: normal App/Chat status and existing Codex observer callers; existing sidebar, cards and contextual help consume the same typed state. No new capability, setting, state vocabulary, execution authority or host-store write.
• Owner: provider-specific Python IO in the existing RolloutCursor; shared activity interpretation and peer routing retain their current owners.
• Related fix(codex): keep Windows rollout reads from blocking native rotation #7222 repairs Windows file-handle sharing/rotation. This changes cursor recovery and shares only the existing cursor test module; it does not replace that work.
• Complete bounded repair proposal. Installed App adoption follows exact-head review and merge; whole-product frontend and sustained host stability are outside this claim.

Validation

• Seven new recovery/checkpoint cases fail against the original implementation; all 114 cursor/host-observation tests pass after the fix. They retain byte-budget, giant incomplete record, source rotation, malformed record, identity and optional-cache checks.
• 23 existing peer-route tests pass.
• Production TypeScript/Chat build passes. Existing packaged goal-activity scene passes with both baseline and changed presentation, including mixed/unknown/unavailable/recovered states, keyboard/Escape help and 390px.
• Real provider readback reproduced the old-cache miss and recovered the native marker using the same bound source and copied cache. Source material and raw operating evidence remain private.
• Ruff and diff checks pass. Diff-scoped semantic advisory adds no vocabulary in this change; risk-based premerge and public-boundary checks run on this exact head. Image manual review is recorded separately from automated validation.
• Main limit: a newest marker beyond the bounded scan or an invalid/unavailable source still reports unknown. No running/idle inference comes from file mtime, Goal history or registration.

Visual evidence

Revision: 8d1983d. Public-safe synthetic packaged browser, 1512×982 default and 390×844 expanded contextual help. Baseline presentation is the PR base; rebase retained the same screenshot source inputs. Actual installed readback is a separate private boundary.

The default view identifies execution uncertainty beside the affected Goal. Mixed live/unknown lanes remain distinct; the narrow help retains each lane’s state, cause and event age in place. No explanation paragraph or control was added. The surrounding populated page is unchanged; this is a status-recovery/readability correction, not a full-page redesign.

Baseline	Changed

Author: model_agent (OpenAI Codex), under maintainer direction. Signed-off-by trailer is present.

[loopx-agent]

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

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

动机

用户打开一个能正常阅读的 Goal，却在侧栏看到状态不可读，无法区分内容故障与执行观测缺失。原来最新回合标记已在文件尾，旧缓存仍让侧栏长期未知；现在同一读取预算内先核对最新尾部，仍缺标记时明确显示执行状态未知。七项旧缓存恢复反例在原实现失败，新实现能读取最新开始、完成或中断标记，未完成记录仍按预算恢复；打包侧栏保留混合活动与真实缺口。本 PR 不改变 Goal 内容访问、执行授权、调度、模型调用或原实验资料，也不承诺所有超长记录一次读完。正常 App 安装采用在评审合并后读回；整个研究页面、其余来源语言和长期稳定验收仍在原任务中继续。

改动思路

同一 provider 的读取顺序与缓存完整性在原 RolloutCursor 内修复，复用有界 Reader 和逆向扫描；UI 仅说明执行观测范围，不新增决策 owner。本 PR 完成交付旧缓存积压下的当前回合观测恢复，以及准确状态文案；原始内容访问、执行和调度权限保持既有边界。
旧路径从已缓存的旧事件继续向前读，大量新增历史会延迟当前状态；新路径只在装载缓存或单次大追加时，用同一个有界读取器核对最新尾部。未完成的巨大记录仍继续既有游标，避免反复重新读取。原身份校验、缓存失败和错误类型保持；无新增权限或决策 owner。这里用文字比较两个分支，未加入未经渲染验收的图。

具体改动

规范依据为 docs/reference/protocols/host-integration-surface-v0.md，规范版本 75f3787。completed host records：implemented，状态取自完整原生记录；8 MiB rollout-read budget：implemented，源校验和恢复共用原有预算；Cached facts are checked：implemented，源检查保留且追齐前不替换检查点。设计规则也要求未知不能推为空闲或执行。

关键代码讲解

• RolloutCursor.observe 校验源与旧事件后，用未读距离以及本次装载／新增量决定是否复用 Backward。限定条件既修旧缓存积压，又保留巨大未结束记录的持续前进。
• RolloutCursor.save 在 forward.pos 尚未到达当前大小时不发布新提示。格式和原子写入保持，源读取不写宿主目录。
• 共享 i18n 将不可读限定为执行状态未知。现有 sidebar、卡片与 ContextHelp 沿用同一类型；既有打包场景覆盖混合、未知、恢复和键盘帮助。四张截图来自同组合成数据，像素、来源和 PNG 元数据均已逐张核对。

对主干的风险

最大反例是没有换行的大结尾记录：若每次预算不足都重新倒查，就会永久恢复不完。当前 head 仅在首次装载或本次新增量超过预算时调整方向；原巨大未完成记录测试和所有宿主测试通过。缓存继续是可丢失提示，不增加共享字段、CLI、任务规则或权限；空绑定、源失败、畸形记录、轮转和陈旧事件保留原含义。真实 SQLite／文件 provider 和实际绑定来源已读回；打包页面的状态为合成 API，不能单独证明宿主执行。Windows writer 与长期稳定不在此资格内，相关 #7222 独立处理。

当前 head 的 137 项宿主／游标／peer 路由测试通过；新七项在原实现全部失败。生产 TypeScript／Chat 构建、现有 goal-activity 场景、Ruff 和语义检查通过。标准 premerge 的6项直接与13项选择检查无失败；原始 ok=false 来自4张二进制图片人工保留项，这4项已另行逐张核对，未冒充自动隐私通过。当前 Goal wait_for_ci=false，未查询、轮询或等待 CI。

我的整体评价

APPROVE，是独立有用且可回滚的增量。长程维度改善持续会话观测恢复，不代表研究收益或生产效率实测提升。用户体验改善来自执行未知与内容可读的区分，真实原因仍可就地查看。实看同一组合成多 Goal、运行／未知／已结束与局部缺口的整页：侧栏把未知限定到执行状态，没有新增说明墙；窄屏弹层完整保留各 lane 原因和事件年龄。视觉证据覆盖这次状态文案和既有弹层；不能认证整个工作台、所有原生宿主或持续稳定性。正常 App 安装采用另行回读。
改动复用已有 provider、缓存和类型，净16行产品逻辑与回归规模合理；更大的索引格式或并行状态 owner 没有必要。未发现阻塞问题。正常安装采用、全页研究体验和持续稳定继续原任务，评审不替代最后准确 head 合并就绪及授权。

English verdict: APPROVE — 8d1983d. Existing bounded observation recovers recent native markers beyond stale cached history, preserves partial/unknown/source boundaries and clarifies execution uncertainty. 137 targeted tests, baseline-sensitive negatives, packaged scene, build and local premerge checks pass; four image holds manually inspected. Installed adoption and general host stability remain separate.