You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Documentation audited against the code. Every documented surface was diffed
mechanically against its source of truth — HTTP routes against llmwiki/web.py,
MCP tools against the operations.py registry, CLI commands against the click
tree, LLMBASE_* variables against their reads — in all three directions:
documented-but-absent, present-but-undocumented, and diverging signature. Ten
misalignments were found and fixed.
Text reads no longer depend on the platform locale.Path.read_text() without encoding= resolves its codec from the locale: UTF-8 on Linux, cp1252 on Windows.
The package writes UTF-8 everywhere — all 38 write_text() calls pass it — but read
it back without specifying one at 31 sites, so kb_stats, llmbase stats and GET /api/stats all failed on Windows against 11 of 226 wiki articles. The bug dated
to April and survived because CI runs on ubuntu-latest only.