TL;DR: llama.cpp b10618 fixes a GBNF grammar parser bug that caused generated tool-call grammars to fail, restoring reliable schema-constrained generation.
Summary: llama.cpp release b10618 (from ggml-org) fixes a parser bug in GBNF grammar handling: parse_char() now accepts '-' as an escaped literal hyphen in character classes. Previously gbnf_escape_char_class() emitted "-" but the parser rejected it, so tool-call grammars failed to parse. The release adds parser and integration tests for hyphens in character classes, with code assistance from Claude Code.
Why it matters: AI builders using llama.cpp for local function calling or grammar-constrained generation should update to b10618 to avoid these parse failures. This fix removes a critical edge case in GBNF tool-call grammars, improving reliability without requiring workarounds.
Source: github_releases