Model roles and the pstack-models config
pstack splits the work into roles. Each role can use a different model, set in one small config file.
An analogy
Like a building crew. Not everyone needs to be the priciest hire. The inspector should come from another firm, so nobody checks their own work.
In pstack
Default split: coding goes to grok. The hardest changes, writing and judgment go to opus 5.5. The review panel has one of each by default. Models from different families reviewing each other is a core pstack practice.
/pstack:setup-pstack checks which models you can use and writes the config. In Cursor it's ~/.cursor/rules/pstack-models.mdc. In Claude Code it's ~/.claude/rules/pstack-models.md. Other tools use ~/.agents/pstack-models.md. Changes apply in a new chat. Claude Code only offers Claude models; see Using it in Claude Code.
Common misconceptions
- "Roles not in the config can't be used": unlisted roles use the skill's built-in defaults. You only change the ones you care about.
- "When the defaults change, my config follows": it doesn't; your old config stays as it is. To get new defaults, delete those lines or the whole file, then run
/pstack:setup-pstackagain.
Try it: Full guide step 2: Pick models and budget (optional)
Related: Tokens and the reasoning budget, Subagents, Host tools (Claude Code, Cursor, Codex)