Fixes a rule that left a broken sentence behind. self-check matched
"double-check your answer before responding" but not the subject and modal in
front of it, so "You MUST double-check your answer before responding." became
"You must." — a sentence that says nothing, in place of one that said
something. Whatever can open one of these instructions is now listed ahead of
the bare form, in both languages.
That bug had been there since the rule shipped. It surfaced within a minute of the feature below existing, which is the argument for the feature.
RuleResult.changes: each rule now reports a short list of what it actually changed, as before/after pairs.hitsstill carries the true total.aggressivehas always come with the advice "read the diff", and the diff was one undifferentiated block for every rule at once — not review, a wall of text with a warning attached. Now an aggressive run is judged rule by rule, and a single rule you disagree with is disabled with--disableinstead of abandoning the level that saves the most.- Empty rather than truncated when a change is too large to summarise. An empty list reads as "nothing to show here"; a truncated one would read as "this is all that happened", which would be a lie.
- Bounded by construction, like everything else that touches untrusted text: the common prefix and suffix are trimmed in linear time, and anything still too large is skipped rather than diffed. Covered by the same adversarial fixtures as the ReDoS suite.
- Shown in the CLI and the web app for aggressive rules, and for every rule
under
--diff. - New public API:
extractChanges,DEFAULT_CHANGE_LIMIT, and theRuleChangetype. - Tests grow from 145 to 164.