v2.0.3

What's New

• Slash-command invocations collapsed to one-line markers. Claude Code stores each /foo invocation as two user messages (the <command-*> tag blob + the expanded skill prompt body). When /search-chat extracted a session that itself contained a slash-command invocation, the consuming agent had to ingest and resist ~100 lines of skill-prompt boilerplate. Both messages are now replaced with a single [SLASH-COMMAND: /name args=ARGS] marker, eliminating the injection surface and context pollution.
• Applied at both index time and display time — new sessions store collapsed content; historical cached sessions benefit on extraction without any cache invalidation.
• Tests updated — 107 tests pass, covering string/list content, args-present/absent, pair drop with skill body removed, preservation of genuine follow-up user messages, and false-positive avoidance when prose merely mentions <command-name>.