When bd is built from a Windows checkout (default core.autocrlf=true), //go:embed reads CRLF bytes from the working tree for internal/templates/agents/defaults/beads-section.md. templateBody()'s LF-only TrimPrefix/TrimSuffix then failed to strip the legacy markers inside the embedded body, so RenderSection wrapped them with fresh markers — producing two nested BEGIN/END pairs in any AGENTS.md written by bd setup mux/claude/etc. bd setup mux --remove only stripped the outer pair (find-first semantics), leaving a stray <!-- END BEADS INTEGRATION --> in the user's file. TestRemoveMuxProjectRemovesBothLayers caught this on Windows. Two complementary fixes: 1. Pin internal/templates/agents/defaults/*.md to eol=lf in .gitattributes. Embedded templates are shipped data, not user-edited text, and should never carry CRLF. 2. Normalize CRLF→LF in templateBody() before stripping markers. Defensive against future templates added without a gitattributes pin. Adds TestRenderSectionFullSingleMarkers as a direct regression guard (asserts exactly one BEGIN and one END in RenderSection output). Fixes #3552
10 lines
505 B
Text
10 lines
505 B
Text
|
|
# Shell scripts must always use LF line endings (especially git hook templates)
|
|
*.sh text eol=lf
|
|
cmd/bd/templates/hooks/* text eol=lf
|
|
|
|
# Embedded markdown templates are shipped data, not user-edited text. Pin LF so
|
|
# Windows checkouts (default core.autocrlf=true) don't bake CRLF into //go:embed
|
|
# bytes, which breaks LF-only marker stripping in templateBody() and produces
|
|
# doubled BEGIN/END BEADS INTEGRATION markers in installed AGENTS.md (#3552).
|
|
internal/templates/agents/defaults/*.md text eol=lf
|