Our agent lessons should be formulated to put more emphasis on action items, not follow up.

We shouldn't record "what was done" to fix this (e.g. in a previous agent session) as a one-off. We should be trying to recommend a change to the agent skills, and/or impose a linting/prek hook-based solution to automatically detect, or a refactor to protect us from the sort of issue again. Usually we would recommend updates to the agent skills.