[loopx-agent]

Goal And Delivered Outcome

Real benchmark display state should drive the research reader. This proposal uses both complete 3 November Grid R7 logs (33/11 records) as the default populated-page reference and an actual saved Graph analysis as the core analysis reference. Synthetic inputs remain controlled boundary/race tests.

Dense original findings reuse the observation/interpretation/measurement reader; absent structured decisions move on demand. A language-specific cache miss reads only the same verified pack's saved unspecified-language original, retaining its identity, citations and usage. The real Graph result also exposed an incorrect headline: one unknown episode hid five recorded utility tests. It now shows 5 / 6 tested, 1 unknown, without claiming positive utility.

Intended target: loopx-project/loopx:main.

Author Declaration

• Written by: model_agent — OpenAI Codex; both commits signed off by LoopX Agent.
• Specification: maintainer direction to replace synthetic core display references with public-safe real benchmark state; existing task-first frontend delivery and trajectory assessment contracts.

Criterion	Disposition / evidence
Recent, complete original display logs	Both R7 arms preserve original text, IDs, times, append order and measurements; independent JSONL byte/hash oracles
Actual saved analysis and costs	Historical Graph: six episodes, all 36 verbatim judgments, original pack/run/revision, 447.0s, 53,768 input / 19,278 output tokens
Negative and unknown evidence	Mixed headline keeps known/unknown distinct; invalid initial result is missing, not zero; native best/retained and final candidate remain separate
Language and source isolation	Current-language cache state stays separate from the saved original; mismatched pack/scope/method/key/language and failed reads do not become accepted evidence
Read-only browsing	Source digests and no-write readbacks; all fixture mutation verbs rejected; production consumer honors execution-disabled responses

Scope And Boundaries

The recent R7 arms both genuinely use benchmark-goal in separate runtimes and are loaded independently. No relationships are inferred from matching questions or adjacent times. Both retain runner_invalid, diagnostic scope and score_countable=false; source-authored confirmation can contain negative or unknown findings. This is frontend evidence, not a causal comparison or qualified benchmark gain.

The older Graph case (11–12 October) supplements those recent logs. It is a six-episode purposive sample, not a third newer whole-history case. Its actual analysis Goal has no Explore log; that gap remains visible. Solver tokens, monetary cost, population coverage and causal effects remain unknown.

The Graph preview replays frozen real GET responses and one public-safe cited native-result range. Original HTTP byte digests are pinned independently. Uncaptured citations fail explicitly. It neither imports the source analysis store nor synthesizes a job to counterfeit real usage. Raw sessions, authority/credential stores, evaluator internals and verifier traces are excluded. Credential/private-path scans found no values to remove from the selected display state.

The synthetic process/SQLite/HTTP test remains separate qualification of the real backend owner. No model/solver call, installed App change, scoring change, release or merge is performed. Broader default/installed adoption and three newer complete cases remain unqualified.

Validation

Current tested source: 1353d017ee5d589ba0c2555f03962b6453cc644a (packaged Chat, real captured-state HTTP reader).

Check	Result / boundary
npm run build:chat	passed; whole dashboard TypeScript and packaged Chat bundle
npm run smoke:research-analysis-read	passed; production consumer negatives, abort/stale reads, failure/recovery, real isolated SQLite/HTTP/job owner with synthetic process output
Python history + fixture suites	11 passed; original five log hashes/bytes, real response byte hashes, language cache misses, citation/unavailable paths, no synthesized jobs, read-only verbs
Packaged research-real-benchmark	passed in 6,717ms; both R7 arms and Graph; all 44 exported original records, all 36 original judgments in zh/en, exact cited excerpt, actual analysis costs, desktop/390px, no writes
Touched Python Ruff + git diff --check	passed
Public/private boundary	16 changed text files clean; supplementary fixture credential/private-path scan clean; attached images visually checked
Manual Ego	packaged captured-state reader; before/after and 390px, no horizontal overflow; 18 existing registered-source file hashes unchanged; one originally absent source path remains absent and zero mutation requests

Earlier c6a70461 also passed the complete research-history and research-record-search regressions and local quick premerge/CQR. Those are historical receipts and are not current-head review or merge readiness. Current-head CI and review remain pending. One bounded rerun of the prior cancelled checkout completed classification, then failed on inherited tests/test_lark_private_conversations.py:62 Ruff E731, with downstream Stage 2C skipped. No CI policy or unrelated test was changed to bypass it.

No RFC pair changes. Vocabulary disposition: the new fields belong only to the local frozen fixture transport; no shared state/protocol vocabulary is introduced. The semantic advisory supports loopx/ carriers and cannot certify the untracked app JSON. Bounded adjacent improvement: reuse existing readers and transport; no new analysis owner.

Frontend / Visual Evidence

Actual packaged UI with real public-safe captured display state. Screenshots are attached, not committed. Reproduce through the reader guide; preview is local only. No installed/live execution readback is claimed.

Graph before: one unknown hides five recorded utility tests.

Graph after: original judgments remain intact; the mixed result is explicit.

390px: saved-original boundary, mixed result, long real questions and negative/unknown judgments.

Recent Grid first-screen and original-record evidence (unchanged presentation)

Before:

After:

Attention review: the first viewport distinguishes native results, measured costs, model judgments and missing history. Complete long repeated questions/reasons remain available. Known results are not erased by one unknown; a utility test is not a positive benefit. The narrow layout wraps without horizontal overflow. Remaining data limits are preserved rather than filled with synthetic success.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · reasoning_effort=max (runtime_reported)

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

动机 / Motivation

回看真实研究过程、读取已保存分析和核对结论依据的工作区用户。
旧页面在中文或英文缓存尚无结果时，把同一分析包已经保存的原文藏了起来；用户只能看到尚无分析，容易重复生成。现在页面直接读取经过身份核验的原文，并保留语言提示、原判断、引用和用量。
已用同一份保存响应比较 base 与 head：旧界面没有原文，新界面可读六个片段的全部 36 条判断；近期 Grid 两组完整展示记录也保持原文、顺序、负面和未知结果。
本次交付限定为研究展示与已保存原文的只读恢复；不重跑模型、不恢复求解器、不证明因果收益，也不代表已安装版本或后续更多真实案例已验收。

改动思路 / Approach

沿用已有 research-analysis reader、分析包和原文引用 owner。当前界面语言已有结果时保持原路径；没有结果才对同一个 pack 发起省略语言的只读 GET，并校验 pack、method、scope 和 result key。readingPack 服务阅读及引用，requestPack 服务当前语言的明确执行、retry 和 cancel，两者分开，避免把历史原文误标成翻译或误用作新执行的身份。

读取身份不变量、已有 provider IO 和派生展示留在各自既有 owner；可选原文读取只观察已保存结果，不生成权限或另建状态，缺失与拒绝保持可见。
本 PR 完成真实记录展示和同包原文可读这一可独立验证、可回退的切片；不承担后续安装推广、更多完整案例或科学采用结论。
普通 finding 直接复用 findingNarrative 和已有完整原文浏览器；没有结构化决定时保留可展开解释，不从措辞、相邻时间或 confirmed 标签推导新状态和因果关系。相比重新生成、另设缓存/翻译服务或只换标签，这条现有 owner 内的修复直接补齐了用户可读结果。

触发	base	head
zh/en 无结果，原文已有保存	显示尚无分析	同包校验后标注已保存原文；不生成
原文身份不一致或其他显式语言	不应替代当前结果	拒绝/不采用；保持错误与作用域
只读响应	后端仍拥有写入权限	前端也尊重本地 explicit false；读取不启用执行

具体改动 / Changes

关键代码讲解

1. readAnalysisSnapshot：首次读取仍按 zh/en 验证完整结果身份；fallback 只检查同一个 pack 的未指定语言原文，不接受另一个显式语言替身。失败不会自动生成分析。
2. useResearchAnalysis：既有 generation/readSequence 在来源、Goal、pack 和语言变化时隔离旧响应；读取原文和执行当前语言仍使用各自身份。readEvidence 按正在阅读的原文 key 请求引用；本地与外部来源的只读动作守卫一致。
3. AnalysisReport：5/6 已做收益检验与1个未知同时可见；“做过检验”没有升级为“收益为正”。六片段全部36条判断、原理由、未知和引用均保留。
4. GoalResearchHistory 让普通记录和 display rewrite 使用同一 finding 解析器，完整原文入口仍在。timeline 的缺失决定说明按需展开，不补造决定或关系。
5. Handler.do_GET 复用真实隔离 HTTP/SQLite 展示 owner；Graph 只返回精确捕获查询，其他引用明确404，所有 mutation405。独立真实 owner/合成 provider 测试与冻结展示证据分开。

按改动前不可变规范 docs/development/design.md，revision 6f90e561ee99a1ef2e3db4e90c94e3c767990c58 审查，逐项为 implemented：Earn The User's Attention（首屏层级和主要测量）；Research Workspace / 研究工作台（原文、时间、来源和未知边界）；Explanations On Demand（完整内容和可达详情）；Populated Page Review（长重复内容、混合状态、窄屏及下方成本）。已先读规范，再审 diff；本 PR 新增的 reference 说明用于披露行为，未拿它自行定义验收。

16文件 +5523/-65，其中4925行来自三份可复用展示夹具。两组 Grid 保留33/11条完整展示事件及44条导出原文/顺序；Graph 是历史六片段定向样本及已保存响应，不冒充更新的完整轨迹。package 注册、既有 browser runner、Python 夹具测试、SSR consumer 与文档围绕同一实际入口，没有另建执行或持久状态 owner。

对主干的风险 / Risks

最需要防的是缓存缺失时串用其他 pack/来源/语言的结论，或把只读原文变成执行许可。已走正反路径：同包原文恢复、错误 scope/key/显式外语拒绝、来源/pack 切换旧响应隔离、本地 explicit readonly、引用404后关闭再读捕获原文。报告4只有8–15行已捕获；24–35行和其他未捕获范围明确不可用，不用虚构内容填补。重载后原文仍可读，四个隔离 reader 的 source_unchanged 均为 true；人工 Graph 路径 mutation_attempts=0。

本次 PASS：11条 Python 测试；base/head 均构建；生产 SSR consumer；real-benchmark browser（7.107s）；原有 history（233.476s）及 record-search（5.584s）；最终 quick premerge 的5项直接检查、3项 catalog 检查和变更路径公共边界扫描。初次两个并发 browser 运行在共享5196端口启动失败，改用各自显式端口后原断言通过；初次 base/head semantic 扫描均缺仓库根 TypeScript dev dependency，隔离 checkout 补齐依赖后 base semantic 与 head quick gate 通过。原失败保留，未将环境失败报成产品通过。

语义与 CI 对齐

新增 Snapshot 字段是已有 Pack 的本地派生展示，复用 typed assessment/语言/只读语义；没有新增共享词汇、持久 schema、peer authority 或控制平面义务。diff advisory 为零支持的 carrier 只说明工具范围，另有 TSX 人工全量 triage 与 full-tree semantic PASS。原生测量、原文、执行默认和新增本地 readonly 守卫均已对照披露。按当前 wait_for_ci=false，未查询、轮询或等待远端 CI，状态 UNVERIFIED。构建原有 chunk-size warning、合成 provider 与冻结 API 边界也保留；没有声称整仓、已安装客户端或真实付费模型资格。

我的整体评价 / Verdict

APPROVE：此 head 完成“真实研究展示和现有原文可读”的有用切片，未发现需要修改的阻塞项。long_horizon=preserved，user_experience=improved；持续科学收益和实际模型采用仍未证明。复用了既有 reader、parser、citation 和权限 owner，局部派生状态与一次条件 GET 的维护成本匹配被复现的问题；没有用样本/测试数替代实际结果判断。

实看了真实记录填充的整页、展开长原文、390px 窄屏和下方方法/成本区。首屏先给原生测量与未知数，判断详情、来源、成本按需展开；失败引用有独立提示，关闭后可读取已捕获引用，原分析仍在。
未检查已安装桌面版本、第三个更新的完整研究案例或科学校准；Graph 是历史六片段定向样本，Grid 是两组近期完整展示日志，均不构成因果或总体覆盖证据。
实看 Ego 视口为1200×888；另外检查了当前 head 的1440×900及390×900自动化截图，包含展开的长原文，另看390×844合成12记录默认页补充布局边界。观察到主数值先行、未知和负面保留，方法/hash/成本位于可达详情；不可用来源有独立提示和关闭路径。

精确审查 head 1353d017ee5d589ba0c2555f03962b6453cc644a；发布前再次核对其未变化。这是作者同账号的 COMMENTED approval conclusion，不是 GitHub formal approval，也不授予 merge/install 或其他审查意见处置权限。

English verdict: APPROVE — 1353d017ee5d589ba0c2555f03962b6453cc644a. No blocking findings; built reader, original evidence, negative/recovery paths and local checks pass. Remote CI, installed delivery and causal benefit remain unverified.