Create a model selection skill

Every task that spawns a subagent or worker has to answer two questions: which model, and how (stay in this session, a subagent, an Orca worker, several models at once). The answer depends on the user, the computer, and the moment, so pstack does not ship a table. This skill writes a skill, select-model, that holds the user's preferences and this computer's facts. The model running a task reads it before it spawns anything and decides. You write it for that reader: an agent that sees it cold, mid-task, on a computer it has never inspected.

1. Find the state

• Target. ~/.agents/skills/select-model/. If the user names another directory, use that one and say so in the final reply.
• select-model exists. Read its SKILL.md, inventory.md, and log.md. This is an update: survey again, show what changed since inventory.md, and ask only about what changed. A re-run rewrites every section of SKILL.md except ## Preferences, which it shows and keeps unless the user changes it. log.md is only appended to.
• It does not exist. A new setup.

2. Survey this computer

Observe, do not assume. Everything below is a command or a file you can read here.