[loopx-agent]

Goal And Delivered Outcome

Research readers need to answer three questions in order: what changed in the task result, what was learned, and which original research records support it? Previously, measurements were hidden in provenance and the expanded analysis led with assessment counts. This proposal makes the ordinary reader follow results → research report → research history, then expands explanations and proposed experiments on demand.

• Outcome basis: the progressive-reading request in this PR, the existing trajectory assessment contract and the frontend/design contracts at 4191b7d966da09a7b0e136353f331e8d5e677dec.
• Observable change: an unanalysed pack exposes existing measurements without a model call; a saved synthesis adds a concise report; Read deeper compares observations with competing explanations and experiments with predictions/disconfirmation. Exact citations and the full history remain reachable with return focus.
• Base: loopx-project/loopx:main. Reuses benchmark-toolkit and its existing typed assessment, process adapter, store and HTTP reader.

Author Declaration

• Written by: model_agent — OpenAI GPT-6 family.

Criterion	Disposition	Owning boundary	Decisive evidence
Results, report, then complete history	implemented	Existing Goal research reader	Packaged DOM order, navigation, source reading and return focus
Useful synthesis beyond counts	implemented	TypeScript trajectory assessment	Same observer call produces grounded findings, alternatives and bounded experiment proposals
Evidence scope, legacy isolation and honest failures	implemented	Qualification/store/process adapter	Named-episode citation checks, unchanged legacy keys/prompts, preserved failure usage/history
Readable desktop and narrow presentation	implemented	Fixed React/HTML renderer	Populated comparison views, 16px main prose, 390px and desktop captures
Independent scientific calibration, user comprehension benefit and installed adoption	out_of_scope	Research/release acceptance	Not established by these presentation or binding checks

Self-check: reviewed the source and applicable task-first/design/public-boundary contracts. The bounded refactor shares citation validation while preserving each metric's narrower episode scope. No new capability, generic framework or configuration owner. The synthetic fixture retains unique end-to-end qualification and readback coverage; it contains no actual research evidence.

Scope And Continuation

New default registrations and CLI preparation explicitly request research_synthesis: true, selecting the versioned research-episodes-synthesis-v1 method. The same call produces the six episode judgments plus at most three findings and two proposed experiments. Every synthesis item names admitted episodes and valid source ranges. The checks bind evidence; they do not prove entailment or causality.

Existing registered configurations retain their exact legacy method, prompt, key and saved results. Setting synthesis false reproduces that method. Reading, expanding, changing report tabs or opening originals does not dispatch a model or run an experiment. The observer produces typed text, never executable HTML/JavaScript. Proposed experiments confer no execution authority. Failed output/qualification attempts retain observed provider usage and are archived before explicit retry; missing observations remain unknown.

Complete within this proposal: generation → qualification → failure/recovery cost → ordinary/deep/source/history reading. Remaining qualification is independent review and maintainer merge, subsequent installed adoption and broader scientific/release acceptance. This author will not self-merge or enable auto-merge. Existing paired roadmap checkpoints retain those distinctions.

Validation

Tested head: 9ca056e7561fb72eb7d0842860320065db5d92ae; source/test commit a0fa7fe35f2b83b138516719be0404ff5339a161. The final commit adds documentation and the captured evidence only. Run state: finished. Inputs: synthetic/public fixtures; isolated disposable SQLite and the real HTTP/source-verification path.

Check	Result	Evidence and limits
Dashboard packaged build	passed	TypeScript and production Chat bundle; existing large-chunk warning remains
Typed assessment contract	passed	25 tests, including synthesis grounding, metric-scope rejection and legacy off parity
Provider/store/real HTTP	passed	56 tests; failed-attempt usage/archive, cache reuse and exact synthesis-only citation authorization
Packaged research-history journey	passed	Native-only ordinary view plus synthesis through real qualification/store/HTTP; results/report/history order, keyboard tabs, source return, history return, 390px and unchanged source/zero mutation attempts; existing scope/language/historical/unavailable regressions
RFC pair and diff	passed	Staged and committed bilingual pair checks; git diff --check
Semantic advisory	passed within tool coverage	No supported new carriers. TSX presentation state and the versioned synthesis extension were reviewed manually; this is not a claim of complete semantic coverage
Premerge canary	automated checks passed; merge held	10 catalog checks, 8 risk-profile checks, public-boundary scan, compile and four direct checks passed. The prior change-quality receipt is stale for this diff; exact-scope requalification and manual visual/benchmark review remain required. No ready-to-merge claim

Frontend / Visual Evidence

The first viewport prioritizes native task results. The ordinary report shows the saved account and expandable findings; deeper reading uses two columns for observation/interpretation and intervention/expected outcome. Narrow screens preserve order vertically. Citations are compact accessible controls; full source text, failure/unknown distinctions and per-episode review are retained. No text is shortened with ellipses to simulate density.

The following are packaged synthetic captures, not live operating context or installed-App evidence. The report fixture deliberately includes a positive later result, an invalid candidate and unknown final retention. Source manifests remain unchanged. A working private preview was checked separately; no public preview service is published.

View	Ordinary: results before report	Deeper report
Desktop, 1280 × 900
Narrow, 390 × 844

The earlier controlled before/after pair uses the same separate native-only fixture at baseline 75f378711 and this branch's original placement fix. It is not a controlled comparison with the new report fixture above.

Native-only first viewport: before and after

Viewport	Before	After
1280 × 900
390 × 844

Whole-page review: desktop makes the result and report entry visible together; narrow reading presents the four native values first. Expanded reports have one active reading focus and keep the per-episode audit secondary. Real long-form reading was checked at both widths; no horizontal overflow. This verifies composition and navigation, not measured human reading speed or scientific usefulness. All published pixels and PNG metadata were inspected for private material. Source-heavy cold-read latency remains an observed experience gap; this proposal does not claim a read-latency improvement or skip source verification.

Type / Area / Direction

Feature and bug fix, focused integration coverage, reference documentation and bilingual roadmap checkpoint. Existing benchmark-toolkit analysis and Goal research presentation. No scoring, task semantics, active experiment, scheduler, permission or authority-store change. No shared-authority migration fixture impact.

Boundary Checklist

• Explicit paths staged; public text and images exclude private state, local paths, internal links, raw trajectories/verifier output and credentials.
• Runtime/default changes and opt-out/legacy behavior are disclosed.
• Paired RFC changes are included together.
• Visual evidence is embedded and synthetic/packaged boundaries are named.
• DCO sign-off on new commits.
• Maintainer review/merge remains separate; no self-merge or auto-merge.

[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)

未发现阻止本次汇报切片合入的缺陷。评审对象为 #72339ca056e。原生测量的四等列、小字和长时间仍是已明确保留的后续修正，不能把此PR称为整体研究工作台优化完成。

动机

研究者打开真实Goal的研究页，需要先知道测得什么、学到什么以及下一步该区分什么。原来只看到历程和过程计数；现在先看到测量，再看到有来源的研究概览，并可直接展开解释和实验条件。保存的汇报、竞争解释和实验建议可分层读取，引用回到精确原文，浏览不会启动模型或实验。原生测量视觉层级、正常App安装采用及科学校准继续验收。不认证模型科学结论，不改变实验评分、预算、调度或权限。

改动思路

既有判读owner负责方法与引用不变量，进程适配器负责执行和失败用量，界面只投影已保存结果；无需另建分析器或调度机制。本PR交付完整的汇报生成、证据绑定、失败保留与分层阅读；原生数值视觉层级和安装采用继续由研究工作台原负责人处理。

我比较了不改、仅前置数字、另建分析器三种路径。前两者都没有回答研究者真正需要的研究结论与竞争解释，独立分析器则会重复配置、来源和调用成本。当前方案沿用冻结轨迹包、同一次明确触发的观察调用、typed qualifier和版本化结果。旧包缺少开关或为false时保留旧方法；新默认注册和CLI准备明确请求新方法。这里的实验建议只是文本提议，没有实验启动权。

入口	读回	权限／失败边界
明确分析所选范围	同一作业生成判读和有来源汇报	失败保留用量；明确重试才归档并重试
打开、展开、换语言或汇报页签	已保存同Goal／范围／语言结果	不启动模型或实验
点击引用	已授权的精确来源行	不能借综合分析扩大单片段指标的来源范围

具体改动

依据是变更前的 docs/development/design.md，版本 4191b7d，而非把本次新增说明当作先前验收依据。Research Workspace 已实现：既有真实Goal研究入口、原生结果、汇报与历史共用原owner，浏览保持读取。Owner language first 已实现：分析语言沿现有配置、缓存、调用和引用路径传递，切换语言不执行模型。Populated whole-page review 已执行：逐一查看9张公开合成图，并重跑真实SQLite／HTTP的打包场景；默认、展开、390px和长混合内容均有观察。指标层级的新修正与安装采用仍保留，未宣称完成。

关键代码讲解

1. prepareTrajectoryAssessment 把显式布尔配置绑定到方法、完整prompt、schema和分析key；不是读页面时临时升级结果。已有语言和展示约束逐句保留。
2. qualifySynthesis 约束最多3项结论、2项实验建议，每项只能引用所列已准入片段的来源及绝对行号。共享的citationsFor也服务原指标，但传入单个episode，故没有扩大其引用边界。
3. _observed_usage / execute 从作业实际事件保留已观测用量。输出解析或qualification失败也写回，不把失败视为免费；未知仍为未知，明确重试继续保留旧尝试。
4. ResearchSynthesis / ResearchReportSummary 使用固定React模板渲染同一结果，普通页先读概览，深入页比较观察／解释及干预／反证。键盘页签、精确引用、关闭和返回历史沿现有controller；生成内容没有可执行HTML或工具调用。

31个路径涵盖typed owner、Python进程／存储适配、固定前端、聚焦集成用例、合成fixture、配对RFC和9张图片。调用搜索核对了base与head及未变更的continuation／citation owner，未增加第二套评估、缓存或调度authority。保留旧格式是为真实持久化key和历史结果，不是无依据增加旧请求解码器。

对主干的风险

最强反例是模型引用了合法但无关的行，或建议了不能区分解释的实验：完整性测试不能排除科学错误。该结果仍属于未校准Agent解释，原始证据及限制可达，没有将其升级为采纳、因果收益或实验执行权限。新注册默认多了综合分析输出，已在参考文档、PR说明和测试中披露；旧配置没有被自动改写或重跑。

独立复跑：typed测试25项、真实进程／存储／HTTP56项、打包research-history场景、前端build和控制平面typecheck均通过。对不可变base和本head使用相同输入，中文、英文和未指定语言的旧模式完整prompt／schema／key／结果逐对象相等；不是只比字段名。失败成本反例在base与head都以同一完整错误拒绝无效输出，base缺少用量，head保留12输入／3输出token，验证了真正修复的差异。越界来源／行、缺少综合分析、未知片段和指标跨片段引用均被拒绝；重试、历史和语言／范围切换由实际owner回归覆盖。

语义与 CI 对齐

这是既有trajectory评估词汇的版本化扩展，局部tab状态不成为控制平面状态。默认关闭／缺省路径的完整legacy对象等价已实测；原生测量前置是另行披露的共享界面变化。实验建议是advisory，启用方法的输出schema及证据绑定是机器强制约束，失败后由现有作业owner恢复。通用Quota／Todo／权限义务没有加入研究专用文案。

标准premerge的直接检查、10项catalog、8项risk和公开文本扫描均通过；原始命令因10个人工项仍返回1，不能称自动门禁通过。我已查看9张合成PNG的像素及元数据，并审查benchmark观察器边界：无真实私有研究、凭证、解题／计分／预算变更或实验启动。RFC双语语义与提交检查通过。语义探针未发现支持的新增carrier，其覆盖有限；方法、配置和可选结果已手工核对。按当前策略未查询、轮询或等待远端CI。独立精确head review、CQR与merge-readiness仍是各自门禁。

我的整体评价

结论为 APPROVE / justified_increment：本次提供了完整可用的汇报阅读切片，规模与现有owner相称；旧持久化分析和后续作业连续性preserved，汇报阅读体验improved。这不是整项研究目标或科学效果的完成证据。

普通页先读原生值，再读概览和结论；深入页将观察与竞争解释、实验条件与反证并排，390px纵向接续，正文16px。完整引用、失败与逐片段核对仍可进入。原生数值的四等列、小字和长时间仍需后续调整，不能据此宣称整页优化完成。

未证明真实模型解释的科学准确性、用户理解速度、正常App安装采用或整项研究验收；原生数值层级继续由原研究工作台负责人修正。

原生数值层级应由原研究工作台owner接续：突出当前捕获窗口最佳有效值，初次和最后值辅助显示，保留结果未知诚实可见，精确时间与来源按需读取。此项是现有显示的明确缺口，未被测试通过或折叠状态掩盖；本评审只批准上述完整报告切片。合并前仍须对未变head取得当前精确scope CQR和ready=true，不以本结论代替安装采用或合并授权。

English verdict: APPROVE — #72339ca056e. The existing observer now generates, qualifies and reads a source-bound research report with explicit execution, preserved failed-attempt usage and legacy request parity. Local25 typed tests,56 real process/store/HTTP tests, packaged desktop/narrow reading and required automated checks pass. Scientific validity, native-value hierarchy follow-up and installed adoption remain distinct; manual image/benchmark review does not replace exact-head merge readiness.