Strong +1 on this. The current global thinkingDefault creates a real tension when you're mixing models with different reasoning profiles in a single agent setup.

The proposed schema is clean. One edge case worth considering: what happens when a user switches to a model that has no thinkingDefault entry and no global default? An explicit fallback order in the spec would prevent silent surprises — e.g. per-model thinkingDefault → global thinkingDefault → "off".

The /think default directive undoing the session-level override is the right call. Session overrides should be transient — if you've configured a model to think, that should be the baseline behavior, not something that gets permanently overridden mid-session.

The revert on model switch behavior is interesting. I'd lean toward re-resolving on /model switch rather than preserving the session override, but I can see the argument for user-intent preservation.

From a practical standpoint: this is exactly the kind of per-model config detail that trips up people setting up multi-agent setups. Having thinkingDefault defined per model in the config from the start — rather than discovered via /think toggling mid-run — is the right ergonomic. Tools like crewclaw.com that generate SOUL.md configs would benefit from being able to expose this per-model field in the visual builder once it lands.