humanspeak / svelte-markdown / 32881417333

95%

main: 95%

push

github

fix(streaming): code fence splitting and marked options pollution (#381)

* fix(streaming): keep unclosed code fences out of the tail-window prefix

An unclosed fenced code block that pauses on a blank line inside the
fence has raw ending in \n\n, so the generic blank-line stability check
froze it into the reused tail-window prefix. Everything streamed
afterwards was lexed in isolation: the rest of the block rendered as a
paragraph and the closing fence became a stray empty code block. This
broke the home-page streaming demo on every run, since its word-chunk
splitter glues the blank line onto the preceding word.

Check code tokens against CLOSED_FENCE_RE before the generic blank-line
test so only a properly closed fence counts as stable. Indented code is
conservatively unstable too, since an indented line after a blank line
continues the block.

* fix(parser): stop marked from polluting caller-held parser options

marked's Lexer writes its default tokenizer back onto the options
object it receives. IncrementalParser reads options.tokenizer to decide
whether tail-window reparsing is safe, so after the first lex any
parser rebuilt from the same object (resetStream, streamId change)
silently disabled the tail window and fell back to a full re-lex per
chunk. In a chat UI that costs every message after the first.

Lex a shallow copy in lexAndClean so caller-held options stay pristine.
The regression test pins both the non-mutation and the tail window
surviving a parser rebuild.

882 of 973 branches covered (90.65%)

Branch coverage included in aggregate %.

2 of 2 new or added lines in 2 files covered. (100.0%)

1457 of 1491 relevant lines covered (97.72%)

5629.41 hits per line