You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Fixed
Installed agents show valid JSON in their output examples (#203, from @rsloveless). The instruction-critic, readability-recommender and corpus-critic templates kept str.format() brace escapes, but none of the three goes through .format(), so a clean install wrote {{ and }} into their JSON schema examples, on both the Claude and Codex targets. An agent that copied its example emitted invalid JSON. The escapes are gone from those three templates, and a new test installs every agent for both targets and parses each fenced json example.
run verify leaves the Sources section out of citation density (#202, from @SID-6921). The citation-density gate counted every [N] and [[...]] in the whole report, including the numbered entries of ## Sources / ## References, so a long source list could carry a thinly cited body over the floor. The gate now counts citations and words in the body only, matching cite-check and the instruction critic. The length, heading and scaffold-leak checks still read the full report.
Sync picks the same winner for duplicate note ids on every filesystem (#201, from @jackiectl2). Sync scanned notes in rglob order, so when two files declared the same id: the one kept depended on directory traversal order; an import of a/note.md and b/note.md could keep b. Sync now scans in lexical path order.