cXTML

cXTML is a dialect of XTML, the markup of the Kimi K3 chat template, for multi-agent prompts on any model.

Origin and credit

XTML (eXtensible Token Markup Language) is the work of Moonshot AI and the Kimi Team, the markup of the Kimi K3 chat template:

• the Kimi K3 technical report, Kimi K3: Open Frontier Intelligence, arXiv 2607.24653, Appendix F;
• the reference encoder encoding_k3.py, published with the Kimi K3 weights (moonshotai/Kimi-K3 on Hugging Face);
• Moonshot AI on GitHub: MoonshotAI, MoonshotAI/Kimi-K3.

XTML replaces the angle brackets of XML with reserved special tokens (open, separator, close) plus an end-of-message stop marker, so every structural boundary is one token. All credit for the notation belongs there. This repository does not redefine XTML; it pins down a dialect. encoding_k3.py is not redistributed here: the codec tests compare against it when you point XTML_ENCODING_K3 at a copy.

What cXTML adds

On Kimi K3 the markers are real tokens. Written into a prompt for any other model they are ordinary characters, and every guarantee that came from being a token is gone. cXTML restores a floor:

1. a quoting rule for markers inside text: code spans and fenced blocks are literal regions; attribute values use character references (grammar/RULES.md 1.6, 1.7);
2. two profiles beyond the original: delegation (task cards between agents, with a strict routing packet) and message (peer messages between agents, call / index correlation);
3. the obligation to validate before delivery to a model without the reserved tokens, with a reference validator and a conformance corpus.

Every document valid under App. F XTML (profile kimi) is valid cXTML.

Layout

path	content
spec/SPEC.md	the specification, GENERATED (rules + grammar + vocabulary)
grammar/	normative rules (RULES.md) and ABNF (cxtml.abnf); App. F vs cXTML origin marked per rule
profiles/	per-profile vocabulary, GENERATED from the validator
validator/	mirror of the upstream reference validator, pinned by SOURCE.lock
corpus/valid, corpus/invalid	conformance cards, one .expect.json verdict per card
validator/codec.py	reference codec (stdlib): one AST, a text renderer and a K3 token renderer, parse for both
tests/	conformance runner (test_conformance.py), codec round trips and differential checks (test_codec.py)
examples/	small annotated documents per profile
tools/	spec renderer, mirror sync check, sanitization gate

Run

python -m pip install regex
python tools/sync_validator.py          # mirror bytes match SOURCE.lock
python tools/render_spec.py --check     # generated files match their sources
python -m unittest discover tests -v    # every card gets its expected verdict
python tools/sanitize_scan.py           # 0 hits

The reference validator is maintained in Russian; its diagnostics are Russian, and the reason substrings in corpus/**/*.expect.json quote them. Exit codes are the language-neutral contract: 0 clean, 1 advisory, 2 malformed.

Status

cXTML/1.0, released 2026-10-03 as v1.0.0. Licenses are a DEFAULT pending the owner's decision: Apache-2.0 for code, CC-BY-4.0 for the specification text (see LICENSE).