# Validation record

## 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 `<style data-plugin="dsh-chat-tidy">` between reads.

| Property | DSH default | With Tidy Tables |
| --- | --- | --- |
| Outer rule | none | 1 px solid `rgba(0, 0, 0, 0.1)` |
| Outer radius | 0 px | 12 px |
| Table minimum width | 0 px | 100 % |
| Header weight | 500 | 600 |
| Header surface | transparent | `rgb(249, 250, 251)` from the host token |
| Cell padding | 10 × 16 px source default | 8 × 12 px computed |
| Wide table | open table width | 999 px scroll width contained by a 746 px shell |

The 8-column table produced internal horizontal overflow while the 1280 px page remained exactly 1280 px wide. At a 700 px viewport the page client and scroll widths both remained 700 px. Under `body[data-ds-dark-theme]`, computed table colors changed to a `rgb(21, 21, 23)` shell, `rgb(44, 44, 46)` header, and `rgba(255, 255, 255, 0.12)` border without a plugin-specific dark-mode rule.

## Visual checks

Captured from the same session, seeded with a reply that exercises headings h1 – h4, paragraphs, nested lists, a blockquote, a table, inline code, a fenced code block, and a horizontal rule:

- desktop at 1280 × 960 in the built-in theme, before and after — [`docs/images/comparison.png`](docs/images/comparison.png);
- the complete reply flow at full height, before and after — [`docs/images/density.png`](docs/images/density.png);
- the whole application window with the plugin applied — [`docs/images/hero.png`](docs/images/hero.png);
- the same stylesheet under the built-in dark theme, before and after, showing the geometry changes while every colour still comes from the `--dsw-*` tokens — [`docs/images/themes.png`](docs/images/themes.png);
- the conversation at 700 × 900 CSS px with no horizontal overflow, where DSH collapses its own sidebar and the user-bubble cap falls back to 88 % — [`docs/images/narrow.png`](docs/images/narrow.png).
- the focused 0.3.0 table component before and after, rendered from the same real assistant reply — [`docs/images/tidy-tables.png`](docs/images/tidy-tables.png).

No plugin error or warning appeared in the browser console.

## Lifecycle check

The Vitest lifecycle test mounts the plugin through a fake client context and asserts exactly one `[data-plugin='dsh-chat-tidy']` stylesheet in `head`, and zero after the effect disposer runs. Because the plugin contributes no preference, no Settings row, and no stored state, disabling or uninstalling it is the complete rollback path — there is nothing left behind to reset.