# Validation record ## 0.4.1 — Issue verification and compatibility Validated on 2026-10-23 with the legacy built Harness checkout `c615b196a2796fce347b2c5c21c9276da7def0ee` and a separately installed official `@deepseek-ai/dsh@0.2.1-rc.2`. All acceptance sessions are synthetic and keyless. The existing user profile and conversations were not used. ### Reproduction - **Issue #4 confirmed:** with 0.4.0 on the published host at a 1280px viewport, the wide shell had `padding-left: 124.5px`, `margin-left: -124.5px`, and a first-cell offset of 125.5px inside its painted frame. The source Markdown had no empty leading column. - **Issue #3 not reproduced as stated:** ordinary paragraphs measured 14px text / 22px line height and 11px between adjacent paragraphs on both hosts. Completed replies stayed visible and work details remained expandable. The report does not include the original conversation, screenshot, or host version, so these checks cannot rule out content-specific cases. The plugin preserves the model’s output language. - **Related current-host compatibility defect confirmed:** 0.4.0 called `project()` on a Session snapshot without `chat`, producing a header-slot `TypeError`. The current host already owns work timing and process disclosure. ### Automated regression evidence Three added regression cases failed on the unmodified 0.4.0 source: inherited wide-table padding and two React subscription/lifecycle cases for a Session without Chat. After the fixes, all 54 tests across 6 files pass. The existing model/DOM suite covers completion, interrupted output, failed/unknown tools, pending interactions, missing history, selection/focus, timers, session choices, scrolling, and teardown. The new style regression also verifies restoration of native padding/margin after disposal. The legacy built-bundle assembled test passes with the actual Harness ModuleLoader, projection, and React renderers. This is separate from browser geometry acceptance. ### Browser acceptance The Codex in-app browser loaded the actual published host and built plugin. After the fix, wide shells have zero left padding/margin and the first cell begins 1px inside the frame (the border). At 1680, 1280, 700, and 390 CSS px, page scroll width equals viewport width, long tables retain internal horizontal overflow, and adjacent paragraph gaps remain 11px. A keyboard ArrowRight scroll changed the long table’s `scrollLeft` from 0 to 10. Short tables and tables nested in lists/blockquotes retain their containment and indentation. Native completed disclosure opens with Enter and leaves the final answer visible. User-stopped history remains accessible. There are no duplicate plugin-owned process controls on the current host and no header-slot error after reload or session switching. The fresh acceptance tab recorded no console error. DeepSeek’s dark theme supplies the shell `rgb(21, 21, 23)` and border `rgba(255, 255, 255, 0.12)` while padding remains zero. On the legacy host, a live keyless turn retained progress while running, its timer continued after reload, and completion changed “Working” to “Worked”, folded intermediate work, and preserved the final Markdown. The current host’s Plugins switch persisted the disabled state but did not unload this installed bundle immediately, including after page reload. Restarting that isolated host removed the stylesheet entirely and restored native wide-table padding (124.5px for the top-level shell; 142.5px in the quote). This is a host activation limitation; users should restart Harness after changing plugin activation. The plugin disposer itself is covered by the lifecycle and style-restoration regressions. `pnpm run check` and `pnpm run pack:check` passed. The dry run includes 35 declared files and excludes source tests, acceptance fixtures, temporary profiles, and credentials. The built-bundle assembled regression passed separately. Reproduction fixtures and instructions live in [integration/](integration/README.md). This record is an acceptance result for those versions and scenarios, not a guarantee for every third-party theme or future host version. ## 0.4.0 — Tidy Work Validated on 2026-09-19 against the built DeepSeek Harness source checkout `c615b196a2796fce347b2c5c21c9276da7def0ee`. This release changes only the standalone plugin; the host was not patched for the feature. ### Automated checks - `pnpm run check` passed, including strict source/test TypeScript checking, ESLint, all unit tests, both bundle builds, and the generated-bundle freshness check. - `pnpm run pack:check` passed; the dry run contained 35 declared release files and excluded private review material, source tests, local profiles, and credentials. - 51 Vitest tests passed across 5 files: turn/final-answer classification, mixed reasoning and Markdown, passive-tool grouping, failed and nested calls, pending interactions, incomplete history, interruption, native DOM mounting, selection/focus protection, manual expansion, session lifecycle, timers, scrolling, disposal, styles, and localization. - Both distributable bundles were rebuilt. The generated-bundle freshness gate is part of `pnpm run check`. - `pnpm run test:harness` passed its assembled regression with the actual Harness ModuleLoader, Chat projection, React renderers, and built plugin bundle. This is a jsdom integration check, not a browser screenshot test. ### Chrome acceptance An isolated local Harness host used synthetic session events and a keyless streaming adapter. All UI interaction and captures used the real browser page. The capture fixture is in [integration/](integration/README.md). - Working header updated with elapsed time; after a page reload it continued from the recorded start instead of resetting. - Progress commentary stayed visible while native reasoning/tool activity grouped into expandable summaries. - Completing the streamed turn automatically collapsed intermediate work and preserved the final Markdown, table, and native response actions. - Expanding work retained its viewport position: the header measured 216 CSS px from the top before and after expansion. - Enter and Space operated disclosure controls. Expanding a group exposed native Think and Bash cards, and the Bash card still opened its command/output details. - Stopping a second turn retained commentary and interrupted reasoning; its work header remained expanded. - English/Chinese controls followed the host language. Light/dark appearance followed host tokens. No warning or error was recorded in the demo page's browser console. Chrome desktop captures used a 1450 CSS px viewport. A separate responsive check used the in-app browser at 700 × 900 CSS px: document client width and scroll width were both 700 px before and after expanding the work. Native sidebar collapse, wrapped commentary, activity controls, and the final table remained usable. The earlier 0.3.0 measurements below are historical evidence for unchanged typography/table styles. ### Publication images [Working](docs/images/tidy-work-working.png), [completed](docs/images/tidy-work-completed.png), [expanded](docs/images/tidy-work-expanded.png), and [dark theme](docs/images/tidy-work-dark.png) are unedited real-app captures. The first three appear on both READMEs and lead the storefront screenshot list. The text, commands, workspace, and results are synthetic; no private conversation, credential, or personal filesystem path appears. The displayed check results are demonstration content, not evidence of a model performing those checks. ## Earlier release: 0.3.0 This record captures the release checks performed for `dsh-chat-tidy` 0.3.0 on 2026-08-30 against a source checkout of DeepSeek Harness `>=0.1.0-rc.6`. ## Automated checks ```sh pnpm run check pnpm run pack:check ``` `pnpm run check` passed strict TypeScript checks, ESLint, 6 Vitest tests across 2 files, both the host and Web-client builds, and the generated-bundle freshness check. The package dry run contained only the declared release files, including the new Tidy Tables comparison image and the eight-entry storefront screenshot manifest. ## Measured browser results The stylesheet was applied to a seeded session in the real assembled Web app — the keyless web lane boots the actual host, a real Chromium, and the built `apps/web/dist`, so the DOM and the application CSS are identical on both sides and only the plugin stylesheet differs. Every row is a computed style read from the live document at 1280 CSS px with a device scale factor of 2. | Item | DSH default | With Tidy Chat | Codex target | | --- | --- | --- | --- | | Body | 16 / 28 px, weight 400, margin 16 / 16 | 14 / 22 px, weight 400, margin 0 / 11 | 14 / 22 px, margin 0 / 11 | | h1 | 24 / 34 px, weight 700, margin 0 / 16 | 24 / 30 px, weight 600, margin 0 / 10 | 24 / 30 px, weight 600 | | h2 | 22 / 32 px, weight 700, margin 32 / 16 | 20 / 25 px, weight 600, margin 20 / 10 | 20 / 25 px, weight 600 | | h3 | 20 / 30 px, weight 700, margin 32 / 16 | 17 / 22 px, weight 600, margin 20 / 10 | 17 / 22 px, weight 600 | | h4 | 16 / 28 px, weight 600, margin 16 / 16 | 17 / 22 px, weight 600, margin 20 / 10 | 17 / 22 px, weight 600 | | List item (not first) | 16 / 28 px, margin-top 6 px | 14 / 22 px, margin-top 8 px | margin-top 8 px | | Table cell | 15 / 25 px | 14 / 22 px | 14 / 22 px | | Blockquote | 16 / 28 px, margin 16 / 0 | 14 / 22 px, margin 0 / 11 | 14 / 22 px | | Rule (`hr`) | margin 32 / 32 | margin 28 / 28 | margin 28 / 28 | | Reading column | 748 px | 748 px | ≈730 px | | **Rendered reply height** | **1411 px** | **1127 px (−20 %)** | — | Every metric landed on its Codex target. The reading column is left alone deliberately: DSH's 748 px is already within 2.5 % of the Codex measurement, so changing it would be churn. Two divergences from Codex are intentional and recorded in [DESIGN.md](DESIGN.md): inline code keeps DSH's `font-size: 0.875em !important`, and the blockquote inset is 18 px rather than 24 px. ## Tidy Tables browser results The 0.3.0 browser lane installed the local release bundle into an isolated DSH Web profile, created a real Workspace and Session through the loopback API, ran the real Agent loop against a local streaming model fixture, and let DSH's production React Markdown renderer produce two tables. Before and after captures came from that one rendered assistant node by disabling only `