[rajanpanth]

Splitting a closed way that a route relation enters and leaves at the same node (a "spoon" loop, e.g. a bus turning loop) throws when the member two places before or after it is not a way:

TypeError: Cannot read properties of undefined (reading 'length')

That is the usual shape of a public transport route, where stop and platform nodes are listed before the ways. With the spoon1 fixture from test/spec/actions/split.js and the members ['s', 'p', '~', '-', '~'] (s and p are nodes), actionSplit('d', ['=']) throws. disabled() doesn't block the split, so the operation is offered and then fails.

Cause

In splitWayMember, the "loops" branch looks one member further (members[i - 2] / members[i + 2]) and passes it to connects(), which reads .nodes. The direct neighbours a few lines above are only compared when type === 'way'; this branch had no such check.

The same branch used i > 2 for the look-behind, so members[0] was never looked at when the split way is the third member. In that case the two halves could be inserted in the wrong order: for ['-', '~', '#', '~'] where ~ and # both run from a to b, splitting # gave ['-', '~', '#', '=', '~'] with # = [a, y], which doesn't touch b where the preceding ~ ends. With one more member in front, the same data was ordered correctly.

Changes

• Only compare members[i ± 2] when it is a way, like the direct neighbours.
• Look behind when i >= 2.

Tests

Three cases added to the "splitting spoon routes" block: a stop node after the ways, stop and platform nodes before the ways, and the third-member case. All three fail on develop (two with the TypeError, one with the wrong order) and pass with this change.

npm run test:once (2305 tests) and tsc pass locally. eslint is clean for the two changed files apart from linebreak-style, which is my Windows checkout.

AI disclosure: this bug was found, and the fix and tests written, with Claude Code. The reproduction and the test results above come from running them locally against develop.

🤖 Generated with Claude Code

[matkoniecz]

close per https://agentscan.tools/user/rajanpanth - classified as large volume account, looks like it submits LLM generated PRs without human review

[matkoniecz]

if it is a real bug: please attach before and after reproduction (screen capture, preferably in form of video) that demonstrates that bug is real and not just LLM hallucination