Setup pstack

Run the /setup-pstack command. It asks for a reasoning budget, then lists the models scoped to this Prime Agent session (enabledModels, the same set /scoped-models shows, or every available model when nothing is scoped) for each role, and writes ~/.prime/agent/pstack/models.json.

If the command is unavailable, write that JSON yourself:

• version: 1
• budget: unlimited (default), large, medium, small, or inherit. The first four map to a thinking= target of max, xhigh, high, medium on rlm.spawn, clamped per model to what that model supports. inherit means children keep the parent thinking level.
• roles: one key per role listed below
• each value is inherit-parent, auto, a provider/id selector, or an array of those
• never write a selector you have not confirmed is available (await rlm.find_models(<query>) returns exact provider/id selectors)
• a role you leave out keeps the shipped roster below

Shipped roster (mirrors the upstream split with models that work here: zai/glm-5.3 takes the fast code seat, anthropic/claude-opus-5 the judgment seat, openai/gpt-6-astra the OpenAI seat, and anthropic/claude-fable-5-1 is the panel's fourth member):

feature, refactoring: zai/glm-5.3
bug-fix: zai/glm-5.3
perf-issue: zai/glm-5.3
hillclimb: zai/glm-5.3
judgment and prose: anthropic/claude-opus-5
hardest tasks: anthropic/claude-opus-5
how explorer: zai/glm-5.3
how explainer: anthropic/claude-opus-5
why investigators: zai/glm-5.3
why synthesizer: anthropic/claude-opus-5
reflect tooling: openai/gpt-6-astra
reflect judgment, divergent, synthesizer: anthropic/claude-opus-5
arena runners: anthropic/claude-opus-5, openai/gpt-6-astra, zai/glm-5.3, anthropic/claude-fable-5-1
arena cross-judge pool: anthropic/claude-opus-5, openai/gpt-6-astra, zai/glm-5.3, anthropic/claude-fable-5-1
swarm workers: zai/glm-5.3
architect runners: anthropic/claude-opus-5, openai/gpt-6-astra, zai/glm-5.3, anthropic/claude-fable-5-1
interrogate reviewers: anthropic/claude-opus-5, openai/gpt-6-astra, zai/glm-5.3, anthropic/claude-fable-5-1

A roster entry with no live credentials in a session is dropped from the injected table for that session (the role then runs on the parent model), so the shipped roster is safe on a machine that has only some of these providers.

Roles: feature, refactoring; bug-fix; perf-issue; hillclimb; judgment and prose; hardest tasks; how explorer; how explainer; why investigators; why synthesizer; reflect tooling; reflect judgment, divergent, synthesizer; arena runners; arena cross-judge pool; swarm workers; architect runners; interrogate reviewers.

Panel roles (arena runners, arena cross-judge pool, architect runners, interrogate reviewers) are arrays: one subagent per entry.

The budget is separate from the model. The upstream plugin bakes effort into model slugs (...-thinking-max); Prime Agent passes thinking= on rlm.spawn instead. The injected table already clamps the budget to each model's ceiling (for example xai/grok-4.6 tops out at xhigh and the opencode-go mirrors at high), so copy the thinking= value as written.

The file is user-level. Do not commit it. If ~/.prime/agent/pstack-models.md exists and the JSON does not, the extension migrates it on session start. A Pi-era ~/.pi/agent/pstack/models.json is not read; copy it across by hand.

After writing, tell the user it applies to new turns. Offer /skill:create-verification-skill once if the project has no verify skill, same as before.