Tokenizer maps now carry the model's tool-calling convention. Auto-derived from the chat template, no per-deployment config.

Tokenizer maps now carry an optional 'tool_calling' block describing how each model packages tool calls (markers, args format, result format), no more out-of-band registry of "qwen2.5 uses '<tool_call>' and JSON args; llama3 uses '<|python_tag|>' and python_args."

'@codecai/maps-cli' (https://www.npmjs.com/package/@codecai/maps-cli) auto-derives the block from each model's 'tokenizer_config.json' chat template; the 70 maps in the codec-maps registry (https://github.com/wdunn001/codec-maps) regenerate with the block populated.

Every polyglot client (TypeScript, Python, Rust, Java, .NET, C) now exposes the field, with full validation that marker names exist as keys in the parent map's 'special_tokens' table.