Configure agent models

Create or update a project-local model and reasoning preference file for skills that coordinate subagents. The file is advisory configuration, not an automatically applied host rule. Only a skill that explicitly reads .agents/agent-models.md uses these preferences. Discover the current consumers from installed skill instructions instead of assuming that every host or installation contains the same set.

1. Detect available choices

Enumerate the models, model families, and reasoning profiles the current host exposes for subagents, including which profiles each model supports. Prefer a host-provided capability listing. If the host cannot enumerate choices, ask the user to provide the supported identifiers and profiles.

Never write an identifier or profile that has not been confirmed in the current environment. The values inherit-parent and auto are always valid. For a model entry, they mean the role should omit both model and reasoning overrides. For the reasoning profile, they mean omit the reasoning override. Never construct a model identifier by adding, removing, or replacing an effort suffix.

2. Load current preferences

Use the repository's documented model configuration when it has one. Otherwise use .agents/agent-models.md in the project root. Read the current file, including its reasoning profile, before proposing changes.

Start a new file with these roles: