Repository navigation
-
Changed: the reply rule. A reply is prose by default. A label line, a list,
or a table is legal when the content has that shape: separate blocks, three
or more parallel items, steps in order, or a comparison across attributes.
Bold stays banned. -
Changed: the Stop hook no longer reports headers and list items. It still
reports bold, em-dashes, slop words, openers, and closers. -
Changed:
evals/run_reply_bench.pydefaults toclaude-sonnet-5-5. -
Tested: the same reply rule ran once through
evals/run_reply_bench.pyon
claude-sonnet-5-5at low effort, 2026-11-10, inside the 2.2.0 rule block.
The replies used lists only for steps, options, or checks, and used no bold
or headers. Two replies used a two-item list, which the rule does not allow.
This release carries the 2.1.1 document rules, so that run is not a test of
this exact prompt. Its files are not committed. No published README figure
covers 2.1.2. -
Known gap:
reader_checkinevals/ste_lint.pystill counts headers and
list items invisible_total, so the reply benchmark scores a fitting list
as a defect. -
Rechecked after the changelog was written: the reply benchmark ran on this
exact prompt against the v2.1.1 prompt,claude-sonnet-5-5, low effort,
2026-11-10, one reply per prompt. No reply used bold or headers, and every
list held three or more items.