Repository navigation
v0.8.13 — say so when a later format flag drops --smart
[0.8.13] - 2026-10-05
Fixed
-
--smartcan no longer be dropped without a word. The output-format flags are one
mutually exclusive family and the last one wins, somcptoon call s t --smart --json
compressed nothing and said nothing: the command exited 0 and printed well-formed JSON,
merely larger. The caller had no way to notice — unlike every other member of the family,
where losing a flag is visible (ask for TOON after--jsonand you can see you got JSON).
The parse loop now reports that one loss on stderr, naming both the consequence and the
fix:mcptoon: '--json' overrides '--smart', so the result is NOT compressed (format flags are last-wins; put '--smart' last) - see 'mcptoon --help'.- Only the
--smartloss is reported, and only when it actually happens. Warning on
every disagreeing pair would nag the order that works —--json --smartis how the
gateway itself calls — about a flag that did exactly what was asked. - Verified against the real CLI, both orders: the note lands on stderr for
--smart --json, and stderr stays empty for--json --smart. - stdout is untouched in both orders. The flags themselves behave exactly as before: the
change replaces sevenelif fmt = ...assignments with one look-up table, and
_effective_fmtstill gives an explicit flag the same precedence over per-tool policy. tests/test_cli_flags.pygains 8 tests — the note's wording, the pairs that must stay
quiet, a literal pin on the format-flag table (a format flag missing from the map would
silently stop selecting a format), and two end-to-end runs throughmain()proving the
note reaches stderr. Suite total: 1909 passed · 2 skipped, so the claim surfaces move
with it.
- Only the
-
Both READMEs and both landing pages now say the same thing about how your agent reaches
your tools. A blind review of the README (one reader, product only, no source) read the
product confidently as "a CLI" — correctly — but that turned out to be because the README
described only one of the two pathsquickstartactually wires.syncwrites a pointer
into the hosts that read one (Claude Code, Codex) and registers mcptoon's ownserve
entry with every host that takes an MCP entry; the README never mentioned the second path,
and one line flatly denied it ("Your agent doesn't connect to MCP servers"). Meanwhile the
landing page had the opposite drift — itsCut:heading framed the two paths as
know which one you're on, i.e. a choice the user has to make, which is the one framing
ADR 0012 explicitly rejected ("the user does not understand the difference between CLI Mode
and Serve Mode, so choosing is not a real option"). Fixed in one pass:- README (both languages): a new "Two ways in — and you don't pick one" block after
quickstart, stating that the CLI is the primary way in, thatserveis registered
alongside it wherever the host takes one, and that the MCP entry ships incompact
exposure (upstream tools stay out of context; reach them via
mcptoon_manifest → inspect → call). - README positioning line rewritten so it no longer claims the agent never touches an
MCP server; it now names both paths and points at the host table for which host takes which. - README takeover line rewritten — it had said takeover "is the mode that actually saves
tokens", which compares two different ledgers (the CLI manifest count and the serve
tools/listcount) and therefore reads as a promise about how much. It now gives the
real, checkable reason: skip takeover and the same tools appear twice, direct and through
the gateway, which costs more rather than less. - Both landing pages: the
Cut:heading changed fromknow which one you're on/
看清你走哪条toboth wired, and the books stay separate/装完都铺好. The paragraph
body was already right (the two ledgers' numbers genuinely don't transfer) and is untouched. CONTEXT.md'sCLI Modeentry said "the only mode mcptoon has" — wording from before
serveexisted, contradicting theServe ModeandTwo Legs Always Onentries in the
same file. Corrected to "the product's defining mode", which is what ADR 0001 actually says.- Verified mechanical: heading swap is 4 lines across the two pages, tag balance unchanged
(7 sections / 6 copy buttons / 5 images each), ZH banned-word law still clean, and the
project's ownmcptoon_gate.pyreports zero new violations on either README (29 and
34 pre-existing, unchanged).
- README (both languages): a new "Two ways in — and you don't pick one" block after
-
The four claims a stranger meets in the first 30 seconds now agree with each other.
A star-curve review found that the surfaces a newcomer reads before running anything —
the GitHub description, the landing page, the README — contradicted themselves. Re-measured
rather than taken on faith; two of the reported contradictions were false positives, one
quoted a value that had already moved, and one real defect was not in the report at all:- GitHub description said
227KB. That was true for 0.8.1; the wheel is now 341KB.
The description is the first line anyone reads and it was nine releases stale. - Both landing pages sold
--toonas "~34% more" on call outputs. 34% is the
manifest (schema) figure; the call-result corpus measures 5.7% (1.8–48.6% by shape,
assets/benchmark_results.json). Quoting a schema number for results is the exact
conflation that got our upstream TOON PR closed — and both pages were missing from
test_no_bare_34_percent_attached_to_toon's file list, which is why it survived. The
guard now covers them. - The README's agent table told you to hand-write Claude Code's and Codex's files.
sync --selfwrites both automatically (measured in an isolated HOME:.claude.json+
~/.claude/CLAUDE.md,~/.codex/AGENTS.md). The row also named the wrong file
(SKILL.mdinstead ofCLAUDE.md). Fixed in both READMEs. - The README said the slim-schema figure is 24% while four other places on the same
page said 88.5%. 24% appears nowhere else; the number was simply wrong.
- GitHub description said
-
Claim surfaces moved to the tree's measured values: 1909 passed (1911 collected − 2
environment skips) and 25,257 physical lines, with the wheel claim at 342KB. The
line figure had been 25,224 — measured before the 0.8.12 help-text edit landed, so it
was 3 lines stale and the guard's 2% tolerance never noticed. Corrected here rather than
left as a known-wrong number; the--smartfix in this same release then moved all three
a second time (1901 → 1909 tests, 25,227 → 25,257 lines, 341KB → 342KB). -
The published wheel-size claim moved with the wheel. The tree is now 350,759 bytes
= 342.54 KB by the guard's own method — normalise the checkout to LF, then really build
the wheel — which is past the 342 line, so the claim moves 341KB → 342KB on all ten
surfaces that state it and 341KB retires. The method was validated against the outgoing
release before being trusted: run on the 0.8.12 tree it yields 349,638 bytes, agreeing
with the 349,540 bytes recorded in that release commit and the 349,666 bytes PyPI serves.
Physical lines moved with it (25,227 → 25,257).