Problem
The update-stack skill was a flat list of steps without a clear mental model. The agent was redundant with CLAUDE.md guardrails and never used organically.
Expected behaviour
Two explicit phases:
- ISO merge — stack modules and lib come out identical to upstream, explicit
--theirs rules per file category
- Project alignment — project modules aligned via MIGRATION.md + tasks reference + ERRORS.md patterns
Agent removed — its responsibilities are covered by CLAUDE.md always-on guardrails.