dc-l1-capetown-waterway-nulls

Analyst notes

Description

Tests whether the agent recognises that a contractor-style GeoJSON can be malformed in three independent ways (geometry is JSON null, geometry is a syntactically valid LineString with empty coordinates, or the required waterway_type column is null) and drops all three buckets while keeping rows whose only oddity is a null name. The prompt deliberately does not enumerate the three defect classes, only says the feature needs a drawable geometry and a proper type, so the agent must inspect the data to discover them and resist the urge to drop every row with any null in any column.

Approach

1. Read the bundled waterways file and inventory the null and empty cases column by column to learn which defects are present.
2. Decide that a feature is unusable when its geometry cannot be drawn or its waterway_type is missing, and that a null name on its own is fine.
3. Filter the features to keep the usable ones, count how many were dropped, and preserve the original attribute and geometry values verbatim on the survivors.
4. Write the cleaned features back out to GeoJSON under the requested filename, and add a top-level dropped_count member on the FeatureCollection so the persona can flag the contractor.

Pitfalls

• Filtering only on geometry.isna() catches the ten JSON-null geometries but leaves the five empty LineStrings and the five surviving null-waterway_type rows in the output, so the cleanup ends up at ten dropped instead of twenty.
• Applying a blanket drop-any-row-with-any-null rule removes the five null-name watercourses that the persona wants to keep.
• Forgetting that pyogrio's GeoJSON driver does not preserve foreign members and writing dropped_count inside a feature property or omitting it entirely.
• Reprojecting the file to a metric CRS before writing, which costs the two CRS subchecks even though the geometric work is correct.
• Filling nulls with sentinel strings like 'unknown' instead of dropping the offending rows, which mutates the attribute values and breaks the per-id attribute check.

Map

Recent runs task v3