runscroll

runscroll — turn one batch run into one scrollable HTML report.

Sprinkle report.add_*() calls through your batch job. Get a single self-contained HTML file out the other side. Mail it, drop it in S3, attach it to a PR. No server. No account. No infrastructure.

When to use this

If your row above says "❌", that other tool is the right fit — runscroll is intentionally narrow.

Install

pip install runscroll                                # core, stdlib only
pip install "runscroll[matplotlib,plotly,pil]"       # with adapters

30-second example

from runscroll import Collector

with Collector("report.html", title="Daily ETL") as report:
    report.add_kv({"started_at": "2026-05-05T09:00", "config": "v17"})

    with report.section("Extract"):
        report.add_text(f"loaded {len(rows):,} rows")
        report.add_table(rows[:5], title="Sample input")

    with report.section("Transform"):
        report.add_text("dropped 142 rows (0.3%)", level="warning")
        report.add_table(dropped[:20], title="Sample dropped rows")

    report.add_text("done", level="success")

That produces report.html — one file, no assets folder, no external CDN. Open it in any browser, mail it, upload it to S3, attach it to a PR.

API surface (the entire library)

Collector(path, title, mode="inline"|"directory", asset_writer=None, log_exceptions=True)
    # context manager: with Collector(...) as report: ...

report.add_text(text, level="info"|"debug"|"warning"|"error"|"success")
report.add_kv(mapping, title="")
report.add_code(code, lang="", title="")
report.add_table(list_of_dicts_or_lists, title="")
report.add_image(bytes_or_path_or_PIL_or_ndarray, caption="", title="")
report.add_figure(matplotlib_or_plotly_figure, title="", description="", close=True)

with report.section(name):              # nested allowed
    ...

That's it. The whole library is one class with eight methods.

Output modes

# inline (default) — one .html file, all assets base64'd in
Collector("report.html", mode="inline")

# directory — index.html + assets/ folder; works as a static site
Collector("report/", mode="directory")

# directory + custom destination — plug in S3 / GCS via AssetWriter
Collector("report/", mode="directory", asset_writer=MyS3Writer(...))

The AssetWriter protocol is one method:

class AssetWriter(Protocol):
    def write(self, relative_path: str, content: bytes) -> None: ...

That's all the library asks. Authentication, region, retries, caching are your concern — runscroll never imports a cloud SDK.

Recipes

Working scripts in examples/ — drop them next to your pipeline as a starting point.

ML training run — examples/ml_training_run.py

Loss curves, confusion matrix, per-class precision/recall, sample worst predictions as inline images. Exercises matplotlib + numpy + PIL + nested sections.

with Collector(out, title=f"Train run {run_id}") as report:
    report.add_kv({"model": "resnet50", "lr": 3e-4, "bs": 64, "seed": 42})
    with report.section("Training"):
        for epoch in range(epochs):
            report.add_text(f"epoch={epoch}  train={tl:.4f}  val={vl:.4f}")
        report.add_figure(plot_loss_curves(history), title="Loss curves")
    with report.section("Holdout"):
        report.add_figure(plot_confusion(y_true, y_pred), title="Confusion")
        report.add_table(per_class_metrics, title="Per-class metrics")

Daily ETL — examples/data_quality_etl.py

Hourly volume, drop-rate warning with a sample of dropped rows, post-clean distribution. The single-file output ships in a mail attachment.

Migration validation — examples/migration_validation.py

Per-table validation with interactive plotly distributions — zoom, pan, hover tooltips, all in the single self-contained file. The plotly bundle is inlined exactly once even when there are dozens of figures. The 6 warnings badge in the top-right corner is generated client-side by counting rs-text-warning entries.

Web crawler — examples/web_scraper.py

Status-code breakdown, per-request latency histogram, every failed URL in a browsable table.

How memory stays flat

Each add_* call serializes its content to disk and flushes immediately. There is no in-memory entry buffer. A 500 MiB report uses the same RAM as a 5 KiB one — only a counter, a section-depth integer, and a file handle live in Python.

This is the design's first-priority guarantee. The test tests/test_streaming_memory.py keeps it honest: 30 × 10 MiB writes must leave less than total_written / 30 resident, and a 30 MiB on-disk image streamed through add_image must not grow RSS by more than 1 MiB.

What this library is NOT

• ❌ A live monitoring dashboard — Grafana / Datadog.
• ❌ A multi-run experiment tracker — MLflow / Weights & Biases.
• ❌ An interactive notebook — Jupyter.
• ❌ A general HTML builder — dominate / yattag.
• ❌ A cloud SDK wrapper — supply your own AssetWriter.
• ❌ A static site generator — Sphinx / mkdocs.

License

MIT.