mirror of
https://github.com/bmad-code-org/BMAD-METHOD.git
synced 2026-11-05 23:53:11 +08:00
Brian
499008ea97
Advanced elicitation now starts only when named and works in the same turn. With no pointer it pushes what the last exchange was about, at its own size; a named target, methods or a direction aim it. With no methods named it runs three pushes from different directions: the method that best fits where the target looks weak, a wildcard chosen as the method whose answer is hardest to predict, and a third push, an open push to reconsider a single choice with everything the conversation holds, or a second method for a section or document. Each method runs as an independent reader with a five-part brief, and the findings merge into one revision of numbered changes tagged with their push, with premise objections shown as decisions for the user. A session is two user turns. The catalog has 48 methods in six directions, 22 recommended, each with a form line the brief quotes; renamed methods resolve by their old names through an aliases column. pick_methods.py takes the resolver output, binds the team's methods_file and rejects an unknown direction or tier. New default_methods setting; the session report records each change as applied or rejected. A new core-tools help topic, help/advanced-elicitation.md, covers how to start and aim the skill, what runs by default, what changed from the earlier version and when a plain push is enough. The reference, help table and changelog describe the rebuild, and the planning skills and agent personas mention it. Trigger queries and four quality eval cases ship in evals/. BREAKING CHANGE: the customize.toml key preferences is now persistent_facts. The catalog columns are num, direction, method_name, description, form, tier, aliases, and the catalog drops from 71 methods to 48. The skill no longer starts on a generic critique request without its name.