[JarbasAl]

Bugs fixed (with evidence)

1. Step pictures were always empty, and could crash the parser — pywikihow/__init__.py::_parse_pictures. wikiHow no longer nests <a class="image"> inside div.step, and no longer sets data-src on the <img> (it now uses src). The old code assumed both:

   • It always found data-src= nowhere in the <img> tag, so HowToStep.picture was None for every step, on every page. Verified live against https://www.wikihow.com/Tie-a-Tie.
   • Worse, on https://www.wikihow.com/Fold-a-Fitted-Sheet (a multi-part article with more <a class="image"> links than steps), the old sequential-index mapping raised IndexError: list index out of range inside _parse_pictures, which was then masked as a generic ParseError by the bare except Exception: raise ParseError — i.e. a real page that should parse fine crashed instead. Fixed by matching each image to its step via the Step-<n> token in its href (the article-global 1-based step number, which lines up with the flat, DOM-ordered self._steps list), and reading img.get("src") (falling back to data-src for forward-compat).

2. Non-ASCII titles left percent-encoded — _parse_title built the title from the URL slug without decoding it. Verified live: https://pt.wikihow.com/Contar-C%C3%A9lulas-no-Google-Planilhas-no-Windows-ou-Mac produced title == "Contar C%C3%A9lulas no Google Planilhas..." instead of "Contar Células no Google Planilhas...". Fixed with urllib.parse.unquote.

3. Search queries weren't URL-encoded — WikiHow.search only replaced spaces with +; any other special character (&, #, non-ASCII, etc.) went into the URL raw, silently truncating or corrupting the query. Fixed with urllib.parse.quote.

4. ParseError discarded the real cause — except Exception as e: raise ParseError dropped e entirely, which is exactly what hid bug Suggestion: iterate over pages until max_results satisfied #1's IndexError as an unhelpful, undebuggable ParseError. Now uses raise ParseError(str(e)) from e.

Tests added (tests/test_pywikihow.py, 13 tests)

• Happy path: title/url/intro/steps parsing, as_dict(), lazy parsing.
• Picture bug regressions: pictures are populated from src, correlate to the correct step (not sequential-index order), and the multi-part fixture that used to raise IndexError now parses cleanly.
• Unicode title regression: %C3%A9 decodes to é.
• Search: encodes special characters, respects max_results, raises UnsupportedLanguage for HowTo/RandomHowTo/search.
• Malformed input: a real "page not found" fixture raises ParseError with the original exception preserved as __cause__.

No network access during tests — pywikihow.get_html is monkeypatched to replay recorded fixtures.

Fixtures recorded (source URLs + when, 2026-08-02/03)

All fetched live, once, via the package's own get_html/requests transport, saved as raw HTML under tests/fixtures/:

• tie_a_tie.html — https://www.wikihow.com/Tie-a-Tie
• fold_fitted_sheet.html — https://www.wikihow.com/Fold-a-Fitted-Sheet
• contar_celulas_pt.html — https://pt.wikihow.com/Contar-C%C3%A9lulas-no-Google-Planilhas-no-Windows-ou-Mac
• search_bitcoin.html — https://www.wikihow.com/wikiHowTo?search=buy%20bitcoin
• not_found.html — https://www.wikihow.com/Special:ThisPageDoesNotExist123456 (404, used for the malformed-input test)

Docs/examples added

• README: documented the previously-undocumented HowToStep.picture field.
• New examples/step_pictures.py showing how to read step pictures.

Packaging

Added a minimal pyproject.toml [build-system] section and a [test] extra in setup.py (pip install -e ".[test]"). Full pyproject-only migration was not done: .github/workflows/release_workflow.yml and publish_stable.yml call the shared TigreGotico/gh-automations reusable workflows with a setup_py: 'setup.py' input and run python setup.py --version / setup.py sdist bdist_wheel directly — removing setup.py would break the release pipeline without also verifying/updating that external reusable workflow, so setup.py stays as the CI-facing entry point.

Model usage

Claude Sonnet 5.