Vally 0.16.0
Released on 2026-09-08.
Breaking
- Grader
namevalues must now be unique kebab-case slugs of at most 60 characters. (#1107) - Public API (
@microsoft/vally):ToolCallEvent.data.argumentscan now be any JSON root value rather than only an object. (#1009) vally-serverschema advances to v8 to distinguish grader execution errors from failing verdicts. Delete existing analytics databases and re-ingest their source JSONL. (#1072)
Features
- LLM grading fails explicitly when the selected evidence resolves to nothing instead of silently treating the result as empty. (#1083)
- Linting now catches invalid directory-discovery config and stops silently skipping evals in broken directory setups. (#1135, #1137)