planning/interview: recommend model + effort per task, researched dynamically

[kyle-sexton]

Idea captured during the dotfiles-agent-drift work (melodic-software/dotfiles#189, docs/topics/dotfiles-agent-drift/PLAN.md deferred items).

The interview/planning flow already assesses task complexity during discovery — it is the natural place to also recommend the session's model and effort level:

• Assess task complexity/ambiguity during the interview and recommend: model tier (capability) and effort level (thoroughness), per the official distinction — upgrade model when the assistant would be confidently wrong despite full context; raise effort when it under-explores/under-verifies.
• When the main model is (or would be) Sonnet, recommend advisor use — Sonnet without the advisor is not the recommended configuration; Sonnet main + Opus advisor is the documented efficiency pairing.
• Guidance must be researched dynamically at run time (docs/model-config, advisor docs), never pinned in the skill — models and recommendations change too often.
• Inverse direction too: mid-task, surface "this is too complex for the current model" when signals warrant.

Sources verified 2026-07-16: https://code.claude.com/docs/en/model-config.md · https://code.claude.com/docs/en/advisor.md · https://claude.com/blog/the-advisor-strategy · https://claude.com/blog/claude-model-and-effort-level-in-claude-code

🤖 Generated with Claude Code