feat(site): read native Explore logs in Time Machine

[loopx-agent]

Goal And Delivered Outcome

A visitor can open an existing native Explore JSONL in Time Machine and read the recorded decisions that changed its direction. The file stays in browser memory: no upload, storage, model call or Agent execution.

Before: prepare a custom capture or search raw records for a decision buried in a summary. After: choose the existing log → read its explicit decisions → filter by recorded outcome → see what changed, stopped and remained → open the exact original. All source records remain one click away. A file without explicit decisions keeps its ordinary reader and an honest empty decision view.

Author Declaration

• Written by: OpenAI Codex, model agent.
• Specification: the visitor journey above; existing Explore history/timeline contracts and docs/development/design.md / docs/development/frontend-delivery.md.

Criterion	Disposition	Evidence
Native file intake without a new format	implemented	Reuses the existing explore_history.ts parser/projection and explore_timeline.ts; shared owners are unchanged.
Find direction changes without inferring them	implemented	Decision view/outcome filter and inline changed/stopped/retained use only the shared owner's decoded decisions. Unknown outcomes remain literal, selectable source values. The public portfolio fixture's 37 decisions (29 replan, 8 continue) each return to their exact original source row.
Preserve complete history and original claims	implemented	All 76 portfolio rows and all 315 curated rows remain readable. Append order, repeated source rows, early negatives, backfills, unknown times and missing links retain their original meaning. Counts describe these fixtures, not research effectiveness.
Failure/recovery and privacy	implemented	Whole-file rejection for malformed/unsupported/conflicting/oversized input; same-file retry, clear, race protection and empty-filter recovery. No stale decision readout. Packaged browser audits observe no file-reading requests or storage writes.

Scope And Continuation

The public-site entry gains one complete local reading journey. CLI and Lark keep their existing log format and behavior; no companion provider, persistent configuration, new shared vocabulary, inference, score or causal graph is introduced. Source labels/outcomes remain claims, and recorded time is not execution duration. Reading neither restores a past state nor verifies a source.

UTF-8, one Goal per file, up to 20 MiB / 20,000 records. Exceeding either limit rejects the entire file. Identical repeated rows remain readable; the existing unique-event projection owns the time strip. Evidence references remain text and are never fetched automatically. The three curated episodes retain their own pinned-capture authority, which local files do not inherit.

Future-facing pass: the local adapter now indexes the existing typed findings once for both selection and rendering, replacing repeated linear lookups. This is a bounded presentation seam, without a second parser or authority owner.

Validation

Tested head: f08dfcccfff1b4c7460e45359d2a85d6b6b8bda1; PR base main. The new decision UI supersedes the earlier flat-reader stage at 35131420bf6c254ed97d10d3d2aafbfde95a4f83; it is not claimed byte-identical to that earlier screenshot approval.

Check kind	Result	Coverage / limit
unit	passed	11 tests: original nine retained plus explicit decision decoding, false mentions/node updates, unknown outcome values, repeated events, composed filters and exact originals. Node 24.21.0.
static	passed	Site TypeScript check and static export; strict documentation and both book builds; diff and bilingual RFC checks. No RFC pair changed.
integration	passed	Risk-based premerge, 4 direct + 5 selected checks. Semantic advisory supports no changed frontend carriers; the local UI reading filters and translation labels were separately assessed. No shared vocabulary changed.
real_entrypoint	passed	The durable three-episode/local-reader browser functions ran unchanged in a scoped packaged harness for EN/ZH × 1440/390. Each view checked all 37 decision-to-original links, 29/8 outcome filtering, complete-record return, malformed/no-decision/unknown recovery, focus and network/storage isolation. This scoped run does not certify unrelated site routes.
real_backend	passed within fixture boundary	Original five-row, 18-hour native-builder fixture qualification is retained with unchanged builder/shared owner. Synthetic negatives and the pinned public portfolio run through actual file input and production TS projection. No active Goal was mutated.
manual	passed	Ego Lite actual chooser/keyboard reading, EN/ZH desktop/narrow screenshots, exact original, empty-filter recovery, whole-history return and isolation.

The full packaged public-site rerun at the tested head passed: four fragment/history matrices (1440/390 × normal/reduced motion), EN/ZH × desktop/narrow Time Machine discovery and complete local-reader journeys, and 190 published entry links. The earlier flat-reader run's unchanged-homepage #learn reload timeout remains a historical failure; this new run used the original assertions and timeouts. Required hosted CI, independent visitor utility, live-Goal and installed-App adoption are not certified. No paid probe, new benchmark job, hosted dispatch, merge, install or release action ran.

Frontend / Visual Evidence

UI impact: changed. The reader screenshots below use the already-public native portfolio records in the packaged static candidate. Before is the flat reader at 35131420bf6c254ed97d10d3d2aafbfde95a4f83; after is the decision reader at the tested head above, with the source replan filter selected. This is proposed source/static delivery, not an installed or deployed App.

View	Before	After
Desktop / 中文
Phone / 中文

English desktop and phone

Entry viewport and malformed-file recovery

The public entry and error treatment are unchanged from the earlier reader stage at 35131420bf6c254ed97d10d3d2aafbfde95a4f83; these historical screenshots retain that boundary. The new packaged journey rechecks chooser, error and retry behavior.

Preview: no branch deployment. Build the public site at /loopx/ and open /loopx/lab/time-machine/?lang=zh, then choose an existing native Explore JSONL.

Attention review: one selected decision and its concrete source changes lead the reading column. The source-derived view/outcome choices support that task; optional search and source/time filters are folded. No graph is invented for a source with zero explicit links. Desktop keeps a quiet source index beside the article; narrow screens put the article first. Raw IDs, JSON and format limits stay behind disclosures. Keyboard focus, unknown/empty/error states, monochrome tokens and reduced motion remain usable.

Type / Area / Direction

• New public-site feature and focused validation; existing Explore display owner.
• Overall roadmap S13 reproducible adoption cases and S11 honest research evidence. Independent-use and research-effectiveness acceptance remain open.
• Shared-authority RFC fixture impact: N/A; no authority, provider or runtime-routing change.

Boundary Checklist

• Public-safe source, fixtures and screenshots; no private state, credentials, raw trajectories or local machine paths.
• Complete local reading proposal; independent CI/review/merge and installed adoption remain distinct.
• No paired RFC changed; bilingual diff check passed.
• First viewport, before/after, responsive and failure evidence appear in this description.
• All commits carry DCO sign-off.

[loopx-agent]

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

Reviewer: actor_kind=model_agent; model=gpt-6.1-sol; provider=OpenAI; reasoning_effort=max; declaration_source=runtime_reported.

精确审查：6742@b9000b2880c0a898a06424e208c9e0d8f9becba1；实际合并基点 c8a01299f2933801b579f58327956dd532f6674b。这是独立代码/本地验证结论；合并权限与 merge readiness 分开。

动机

已有原生 Explore JSONL 文件的访客，需要找到研究为什么改变方向，并回到那条记录的完整原文。 原页面只展示三个策展案例，自己的日志需要手工加工或逐行找决策；候选让访客选择原文件后直接阅读明确记载的决策，再按结果筛选或查看所有来源行。 实际静态构建中，键盘选文件后直接出现决策；五条真实原生构造器记录保留18小时跨度、负面结果、回填和未知关联。公开捕获的37条决策逐条对应原始来源行，29条调整方向与8条继续可筛选；错误文件整份拒绝，换文件和返回案例均恢复到正确上下文。 这是网站本地阅读交付；不上传或持久保存文件，不恢复历史状态、不执行 Agent，不证明研究收益、完整站点资格或已经部署上线。

改动思路

现有 TypeScript Explore 编解码与时间投影继续负责记录语义，新增代码只承担浏览器文件读取、局部阅读筛选和展示，不创建第二套研究状态或评判规则。 当前 PR 交付原生文件到明确决策及完整原文的实际网站旅程，保留原有三个案例；没有新 CLI、Lark、模型或执行配置。 维持现状要求访客自行加工文件；只给命令或下载链接仍不能完成阅读。复用现有 Explore owner 加一个本地文件适配器，比新增上传服务、转换格式或独立状态机更小。

路径	本头可观察结果
选择原生文件	直接读来源明确记载的决策；原文和全部记录仍可达
无效或切换文件	错误整份拒绝；新文件保留，旧读取不能覆盖

此表来自已检查的源码和下述执行边界；不表示安装、模型理解或部署证据。

具体改动

独立依据是 docs/development/design.md，不可变版本 8795fc042550c875d78b8bfc1a4874523fbb6693。Earn The User's Attention：决策先读、搜索选项折叠、原文可达；Accessibility：键盘选文件、恢复焦点和44px控制；Responsive Behavior：390px阅读先于索引且无横向溢出。三项在本轮构建与实际浏览器旅程均已实现。

新增适配器保留全部来源行，单独复用 Explore 的去重投影；时间条表示 recorded_at，不推断执行耗时或因果。只识别既有解码器给出的明确 Outcome，未知值保持原文，不用关键词猜测决策。默认有决策就直接读决策，无决策时读全部记录；两个视图、来源标记、时间和文本筛选都不改源顺序。

文件版本键与读取清理防止旧异步结果覆盖新文件。同一文件可重试，取消不清掉已有阅读，退出恢复策展上下文。原文用字面文本渲染，证据引用不变成自动打开的链接。20MiB/20,000条为明确整份拒绝边界；50条分页不截断数据。旧3513平铺阅读图片是历史，不能代替本次f08决策旅程验证。

关键代码讲解

• readLocalExploreLog：Strict whole-file intake reuses Explore validation/projection and preserves source lines.
• selectLocalLogEntries：Local exact outcome/status/time/search filters preserve append order and unknown values.
• LocalLogReader：Async file state and unmount guard prevent a stale previous read from winning.
• LoadedLog：Explicit decisions, original text, pagination and recovery use the same source entry.
• TimeMachine file input：One file choice starts local reading; cancel preserves and exit restores the existing case context.

对主干的风险

未发现阻塞问题。真实浏览器确认坏JSON第2行、无效UTF-8、混合Goal和冲突事件整份拒绝，重复行仍独立可读；来源HTML保持字面文本，未知决策标签不被重解释。受控延迟旧 File.arrayBuffer 后，后选文件仍保持37/76；这里的延迟是浏览器边界故障注入，不是伪造后端执行。文件操作期间fetch/XHR/beacon、存储和新资源请求均为空。

同一不可变基线和候选构建的三个策展案例，问题、结果、原文及来源完全一致；网站静态包实际运行，native构造器和append owner实际产出测试日志，没有用手填UI响应代替真实读路径。

本轮验证：当前 head b9000b2880c0a898a06424e208c9e0d8f9becba1 是合并 main 的提交。实际浏览器执行绑定 f08dfcccfff1b4c7460e45359d2a85d6b6b8bda1，这段源代码、共享解码器、原生生成器和设计依据均未改变；新头重新构建的119个静态文件逐字节相同，故复用该实际浏览器证据，不把旧执行重命名为新头。当前头11项测试、构建、前缀包、语义advisory及标准premerge已另行通过。第一次误用不存在的npm测试别名未运行测试，改用仓库实际node测试命令后通过，失败诊断保留。11项Node测试全过，TypeScript/Vite/SSR及前缀静态构建通过；标准premerge的4项直接检查、4项目录检查和1项边界检查全过。实际Ego桌面1440与手机390布局、中英文、减少动态效果、无横向溢出和至少44px新控制均通过。初始评审fixture少传expected_goal_id已在生成前修正；浏览器文本选择器错误、长requestAnimationFrame循环超时及远处分页指针可见性失败均保留，随后在同一Page清理待执行脚本并用正确DOM等待和真实键盘操作读回，未弱化断言。完整Playwright站点目录未独立重跑；作者早期home #learn超时仍是历史全站资格未知，不声称已修复。没有部署、已安装App、付费模型或研究收益证据。

我的整体评价

APPROVE：本次有界原生文件阅读目标达到，long_horizon 与 user_experience 均 improved；旧来源、负面结果与决策理由可以带入后续阅读，所需额外输入只有本人选文件这一项缺失本地事实，没有转换、配置或重复确认轮次。隐私取舍是刷新后需重新选择文件，界面如实说明。局部UI词汇复用现有类型化 owner，已做的Map/解码器复用就是本次合适的前瞻收敛，不需再造通用框架。 现有 TypeScript Explore 编解码与时间投影继续负责记录语义，新增代码只承担浏览器文件读取、局部阅读筛选和展示，不创建第二套研究状态或评判规则。 当前 PR 交付原生文件到明确决策及完整原文的实际网站旅程，保留原有三个案例；没有新 CLI、Lark、模型或执行配置。 本次没有执行合并、升级浏览器或更改已安装产品；没有查询、轮询或等待远端CI。

English verdict: APPROVE. The built native-log journey preserves source lines and explicit decisions, rejects invalid files as a whole, handles replacement/recovery, and retains the three curated cases. Local checks and the affected real browser paths passed; full-site, deployment and research-utility qualifications remain separate.