[loopx-agent]

Goal and delivered outcome

Research reports placed version/source bookkeeping before the saved account, and a long overview pushed “Next experiments” below the first mobile viewport. This change collapses version details, retains direct access to the previous report and unmatched-comparison warnings, and puts the findings/experiments tabs ahead of the overview. Selecting experiments starts with the experiment; returning restores the full saved overview.

用户可以先选择读结论还是下一步实验，再按需查看版本变化、研究片段和原文。完整正文、范围、未知与失败状态保留。

Author declaration

Written by: model_agent — OpenAI GPT-6 family through Codex.

Implemented against docs/development/design.md (Research Workspace, Explanations On Demand, Populated Page Review) and docs/development/frontend-delivery.md (Task-first delivery), at base 1b1ae8e5685e443a4585c9e58067df893e710290.

Criterion	Disposition	Implementation / evidence
Default view prioritizes the requested work	implemented	Collapsed version details; report reading tabs before the complete overview
Requested detail remains readable and recoverable	implemented	Full-width version detail at 390px; previous version, Back, citations and exact focus return
Packaged journey with actual owner and negative cases	implemented	Native HTTP/SQLite report fixture, read failures, language/keyboard/narrow reading; qualification limits below
Public first-screen evidence	implemented	Same synthetic reports at desktop and 390px, before/after below

Scope and continuation

Complete within this presentation slice. It reuses the existing benchmark-toolkit analysis owner and saved report. No analysis input, cached result, citation identity, scoring or execution authority changes. The adjacent cleanup removes duplicated version counters and presents provenance as label/value rows. No new module, protocol vocabulary or settings surface.

Existing comparison semantics and the fixture correction in #7304 remain with that proposal; they are not duplicated here. Installed adoption, independent comprehension and scientific benefit remain unqualified. Maintainer review/merge is separate; auto-merge is not enabled.

Validation

Tested revision: 68c627f. Run state: finished. Input classes: synthetic, public_fixture, authorized_private_read_only.

Check kind	Result	Evidence / limitation
static	passed	Dashboard typecheck/build, packaged chat build, CSS custom-property check and diff checks
real_backend	passed	Disposable native HTTP/SQLite owner reads synthetic saved analyses; model provider is a deterministic test process
integration	passed	Complete packaged research-history scenario, including report tabs, complete overview restoration, optional version detail, previous/Back, citations, language and keyboard/390px readback; uses the already-proposed fixture compatibility change from #7304 at b101f39952f6c4a9416241e77cc5745e95f033f7
integration	failed	Unmodified base fixture inherits synthesis revision 2 while disabling synthesis; native validation correctly rejects it. This known fixture failure is retained, not reported as a stock green run
manual	passed	Two existing long reports, unknown-pack error/Back recovery, citation output and exact focus return; read-only inputs/results and saved failures/usage remain unchanged
static	passed	Standard premerge: 4 direct + 8 selected checks, no failures; 5 image-review holds remain
manual	blocked	Independent visual review of the five public captures and final maintainer approval remain outstanding

A post-rebase stale-bundle startup was rebuilt before the final exact-head browser run. No live model or benchmark was dispatched. This validates reading and preservation, not a user-study improvement or installed release.

Frontend / visual evidence

UI impact: changed. Source data: synthetic. Packaged source UI, not an installed App. Before: bd038d4f3ca9ec22b06df112c151f6458d2653d3 (the affected baseline renderers match the current base). After: 68c627f. The same reports include a changed configuration, negative/unknown results and competing explanations.

The default viewport replaces repeated bookkeeping with one version disclosure; the comparison warning and previous-version control stay visible. Tabs remain reachable before a long overview. Requested provenance uses body-size label/value rows across the full narrow width. Existing tokens and status treatments are retained. No body text was shortened or reduced in size.

Desktop before:

Desktop after:

390px before:

390px after:

Next experiment, one tab away:

Boundary checklist

• Public-safe source and synthetic images; no private state, raw evidence, credentials or machine paths.
• No duplicate benchmark work or scoring/runner changes.
• Scoped to report reading; existing semantic owners retained.
• No paired RFC changes; bilingual check passed.
• Responsive first-screen evidence included.
• Signed-off commit; proposal left for maintainer review and merge.