When updating diagram parsers (Jison/Langium/Chevrotain), treat lexer/tokenization precedence as a compatibility contract: ensure changes don’t break existing inputs and that intended tokens are unambiguously chosen. Apply these rules: 1) Assume the first/earliest matching token wins (Langium/Chevrotain behavior). Avoid overlapping regex/patterns that allow one token to “shadow” others. If overlap is unavoidable, control it explicitly (tighten patterns, reorder tokens, or use a custom matcher that refuses matches that should belong to other token types). 2) Never change the meaning of a previously-supported character/string without explicitly preserving the old interpretation (or providing a migration plan). If a symbol like `-` previously parsed as a generic/node string, keep that option working while adding the new operator token. 3) If you remove/replace lexer terminals to resolve conflicts, confirm that downstream validation preserves the same semantics the feature provided (don’t silently drop behavior). 4) Add regression tests for prior real-world diagrams/labels that previously worked—especially cases involving punctuation, quotes, and ambiguous terminals. Example (conceptual fix for “shadowing”): - Instead of defining a broad token that can match everything, define a narrower token or use a matcher that checks the input context before returning the token. In code review, ask: “What token now matches this substring, and does that substring still parse the same way as before? Which token wins, and why?”