microlinkhq / html-get / 40592474037

97%

push

github

perf: lazy pandoc probe, bounded content-type cache, one-walk url rewrite (#303)

* perf(pandoc): probe the default pandoc lazily and async

getHTML evaluated defaultPandoc() as a default parameter, so the first call of every process ran execFileSync('pandoc --list-input-formats') even for plain HTML. That blocked the event loop: synchronous part of the first getHTML call 33-47ms (5 fresh processes), now 7-10ms. The probe now runs only when a response is an office document, through tinyspawn, once per process. Explicit `pandoc` values and the exported sync defaultPandoc() behave as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(util): bound the content-type cache

The cache was keyed by the raw remote header and never evicted: 200k distinct headers retained 64.6MB. It now resets after 1000 entries. Hit path unchanged (15.7 vs 16.6ns per two lookups, median of 5).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* perf(html): rewrite url attributes in one tree walk

rewriteHtmlUrls ran one full-document selector query per url attribute (10 walks). It now selects every url-bearing tag once and checks that tag's attributes. Output is identical: an element removed for a local address is dropped either way, and surviving elements get the same rewrites (checked on every test fixture plus crafted nested local-address cases, 48 renders). html rewriteUrls on: 143.0 -> 111.2ms (-22.2%, 6 alternating runs per side).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* test: lock lazy pandoc, cache eviction, and url rewrite

The new branches still passed when inverted. Assert the cache cap, one shared probe, and url attributes that share an element.

Co-authored-by: Cursor <cursoragent@cursor.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

646 of 712 branches covered (90.73%)

Branch coverage included in aggregate %.

73 of 73 new or added lines in 3 files covered. (100.0%)

2020 of 2051 relevant lines covered (98.49%)

162.92 hits per line