The reason check split on white space and trimmed punctuation before
looking a phrase up, so a stored name ending in punctuation could be
trimmed into a different Regular (MQ-100-4) and a name joined to the
next word by punctuation was never seen (MQ-100-5).
A mention of a Regular is now where the reason spells its name, ignoring
case and spacing, with the name's own outer punctuation optional and no
letter or digit just outside it. A mention names its Regular unless a
different Regular's mention covers the same text or more, which keeps a
mention of "Easy Beef Chili" from naming "Beef Chili" (MQ-100-3).
One table-driven test covers the whole matrix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The reason check compared normalized names, so a reason naming only
"Beef Chili" passed for a citation of "Easy Beef Chili". Find the
Regulars the reason names with the same find the citations use, taking
the longest run of words at each point, and require a cited Regular
among them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A cited Regular now resolves by exact name (ignoring case and spacing)
first, then by normalized name only when that names a single Regular;
an ambiguous citation drops the Side quest instead of rewriting it to
whichever Regular shared the key (MQ-100-1).
A Side quest name with no words left after normalizing is malformed and
goes through the repair turn, so it can no longer evade the dedupe or
the dinner-name rule (MQ-100-2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The skill asks for twice the Side quests needed, using the household
context plus the pending Side quests, through an embedded, checksummed
prompt template. A malformed reply gets the harness's one repair turn.
A well-formed Side quest is dropped (and listed in Dropped) when it
repeats a Regular, Quest log dinner, pending or passed name or an
earlier Side quest in the batch, cites no Regular or one that is not a
Regular, or has a reason over 180 characters or naming none of its
cited Regulars. Golden tests use llm.Fake.
Maestro task 85a898e1-f1f5-4414-92f3-3277d86afe90.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dev_agentandClaude Opus 5.5