Put report_tools blocks and table rows on their own lines (ENG-3358)

[StpMax]

Why

Pages built with report_tools put everything after </style> on one line: a 9 KB dashboard had a 6,558-character line holding the whole visible page. Line-based tools get nothing useful from that: read_text_file cannot read one section by line range (and shortens a very long line), html_lint reports every console error on the same line, and a diff shows the whole page as one changed line.

Result: each block and each table row, list item, chart bar and select option is on its own line. The rendered page is pixel-identical to before. rt.update / rt.insert keep this layout and still work on existing single-line pages.

Changes

• Block helpers (section, para, bullets, table, bar_chart, filter_table, data) end their markup with \n; page() puts head tags, <body>, <main> and the closing tags on their own lines.
• Containers still join parts with "": a line break between two inline parts (text and a link in a section) would render as a space. For the same reason the filter's </label><select> stay together.
• update / insert keep the layout:
  • a kept section heading keeps its trailing line break, so an updated section is byte-identical to a freshly built one;
  • replacing a whole element does not double the line break that follows it;
  • insert(after=) puts an Html block after the element's line break; inline content and plain text stay right next to the element.
• New markup (including the filter script) uses the page's line break, taken from its first one (a \r\n inside a shown value does not count).
• update with rt.table(..., id=) now replaces the table's table-wrap div too. Before, every such update nested one more wrapper (pre-existing bug).
• table and filter_table share one _table_wrap helper.

flowchart LR
  A["rt.para / rt.table / ...<br/>markup + '\n'"] --> B["rt.section / rt.page<br/>join parts with ''"]
  B --> C["one block per line<br/>no breaks inside text"]
  C --> D["rt.update / rt.insert<br/>keep line breaks around edited elements"]

Loading

Example of the new layout:

<section id="detail-section"><h2>Detail</h2>
<div class="table-wrap"><table id="d"><caption>Cap</caption>
<thead><tr><th scope="col">I</th><th scope="col" class=num>G</th></tr></thead>
<tbody>
<tr><td>A-100</td><td class=num>40</td></tr>
<tr><td>B-200</td><td class=num>0</td></tr>
</tbody></table></div>
</section>

Existing pages stay single-line until rebuilt; nothing migrates them.

Testing

• tests/test_report_tools.py: 41 passed. Six tests that pinned exact markup were updated. New tests check that:
  • every line break in the body sits between two tags;
  • rows, bars and options are one per line;
  • update + insert produce the same bytes as rebuilding the page;
  • inline and plain-text inserts stay next to their anchor;
  • edits on a CRLF page add no bare \n, and a CRLF inside text does not make new markup CRLF;
  • repeated update of a table yields the same page as a rebuild (one wrapper).
• Rendering: the same page was built with the old and the new module and edited with the same update / insert calls. Both were rendered in headless Electron in light and dark themes, at 1200 px and 375 px, with three filter selections. All 12 bitmaps are identical. A control run with one extra line break between </label> and <select> was reported as different in all 6 affected cases.

Evals (cowork_evals, quick check)

baseline = origin/staging 2727d049, eng-3358 = this branch 925aa781; same cowork-server (2fe11f41), mindshub_air, LLM judge sonnet, 2 repeats per variant. The variants differ only in report_tools.py and its tests. Two repeats are a smoke check for regressions, not a statistical comparison.

Task	baseline judge	eng-3358 judge
dialog-dashboard-refine (5 turns, built with rt.page / rt.section / rt.bar_chart, extended with rt.insert)	0.62, 0.12	0.86, 0.74
dialog-incremental-report-01 (control: builds a Python script, does not use report_tools)	0.15, 0.25	0.96, 0.15

• dialog-dashboard-refine: all four repeats produced a dashboard. The baseline 0.12 is a wrong final answer (Europe instead of APAC), unrelated to report_tools. Program checks pass in 3 of 4 repeats (the failure is that same baseline answer).
• Layout of the real artifacts: baseline dashboards have 14-15 lines, the longest 5,160-5,332 characters. The eng-3358 dashboard built with report_tools has 292 lines, the longest 288 characters. In the other eng-3358 repeat the agent wrote its own SVG chart, which keeps its own long line.
• Errors in agent turns: both variants hit the same pre-existing rt.check error ("HTML tags shown as text" after passing raw HTML as a plain string) and recovered. No new errors from rt.insert / rt.check.
• dialog-incremental-report-01 spread is the known permission-question noise of that task.

🤖 Generated with Claude Code