activeing123/mcptoon v0.8.6

on GitHub

This release is broken. The published wheel is missing mcptoon/_toon/__init__.py, so mcptoon --version fails with an ImportError. Upgrade to 0.8.7.

v0.8.6 — 2026-10-02

Fixed

• The TOON encoder was three releases out of date, and the spec document said it
  wasn't. mcptoon vendored python-toon v0.1.1 (2025-10-30) while upstream shipped
  v0.2.0 (2026-09-29) whose release note reads "conform to TOON spec 4.1" — and
  docs/TOON_SPEC.md already claimed spec v4.1 alignment. The document described a
  conformance the shipped encoder did not have. This is what the toon maintainer meant
  when they closed our PR with "the TOON encoder it bundles is long out of date".
  0.1.1 emitted the spec's retired comma length marker (rows[3,]{id,n}:) and, in
  mixed arrays, dropped the list-item prefix on nested arrays — so {"mixed":[1,{"a":1},[2],"str"]}
  did not survive its own round-trip. Measured over 320 payloads: 0.2.0 round-trips
  320/320, 0.1.1 only 319/320, and 77.2% of outputs are byte-identical — every one of
  the 73 differences is 0.1.1 producing non-conformant text. Output is never materially
  larger (1/320 grew by one token; net −2.76%).
• --toon help text and the README quoted the wrong number. "~34%" was measured on
  tool schemas; --toon encodes call results. Measured on both payloads:
  schemas 11.8%, call results 5.7% mixed and 1.8%–48.6% by result shape (a list of
  records barely shrinks; a single record halves). 18 doc sites corrected.
• The 34% was not reproducible by anyone, including us. The script that produced it
  is gone (bench_tokens.py emits json/slim/compact but never had a TOON row) and the
  corpus was never committed, so the encoder moving 0.1.1 → 0.2.0 went unnoticed.
  scripts/make_benchmark.py now carries its corpus, seed and generator, and
  tests/test_benchmark_repro.py (12 tests) pins the assets to a fresh run, refuses to
  let the historical table be overwritten, and fails if any shipped doc attaches a flat
  percentage to --toon. It caught three missed sites the day it was written.
• The footprint guard could not see subpackages. SRC.glob("*.py") is not
  recursive, so _toon/ — the repo's first .py subpackage — was invisible to it. Now
  recursive: 24,148 lines across 43 modules.
• --toon no longer errors on input the old encoder accepted. Decode is
  non-strict by default (upstream ships strict=True); empty input returns {}.

Changed

• Wheel 308KB → 322KB (329,878 bytes measured). The growth is packaging, not bloat:
  upstream's package layout plus a decoder that is 23.7KB of source. 308KB is retired
  from every live surface; CHANGELOG.md keeps saying 308KB because that is what 0.8.5
  actually shipped. 23,485 lines / 34 modules → 24,148 / 43.
• ruff excludes src/mcptoon/_toon/. Vendored code is not ours to reformat; the
  upgrade path stays a straight replacement. Our wrapper is linted normally.

Notes for downstreams

• TOON text produced by 0.8.5 and earlier in the retired [N,] form cannot be decoded
  by 0.8.6. That form was never spec-valid, so this only affects output that was
  already unreadable. Text produced by any spec-conformant encoder is unaffected.
• VENDORED_VERSION and TOON_SPEC_VERSION are now exported from
  mcptoon.toon_vendored, so the shipped version can be read instead of guessed from a
  doc string.