Ps Setup Pstack

Set up P-Stack

Configure one repository for the user's selected agent harnesses. The generated model registry is shared. Harness-specific files are adapters derived from it.

Read the harness reference before asking questions or writing files. Run scripts/configure.py --help before using the renderer.

1. Choose harnesses

Ask which harnesses the user wants to configure. The supported values are pi, codex, cursor, opencode, and claude.

Do not create .pi/APPEND_SYSTEM.md, a Codex instruction import, Cursor rules, or OpenCode instruction entries. Pi, Codex, Cursor, and OpenCode read root AGENTS.md directly. Claude Code is the only listed harness that needs a root instruction adapter.

2. Resolve model tiers

For each selected harness, ask for three model choices:

• small for fast, bounded work.
• medium for ordinary implementation and investigation.
• large for design, judgment, and difficult work.

For every tier, collect the user's display label, the exact harness model identifier, and the exact reasoning level or variant. Treat the user's label as intent, not as an identifier. Use the active harness catalog and current primary documentation to resolve it.

For example, a Codex user may choose GPT-5.6 Luna, Terra, and Sol. Resolve and verify the exact slugs exposed to that account. Do not assume an example slug is still current.

3. Verify before writing

Probe every model and reasoning pair with a small read-only request. Save how each pair was verified and the current date in the plan. If a harness can prove only that it accepted the requested pair, say that. Do not claim that the provider exposed hidden applied reasoning.

A failed or ambiguous pair stops setup for that harness. Never substitute a fallback silently.

4. Build the plan

Write a temporary JSON plan with this shape:

{
  "schema_version": 1,
  "harnesses": {
    "codex": {
      "harness_version": "version reported by the installed harness",
      "tiers": {
        "small": {
          "label": "GPT-5.6 Luna",
          "model": "exact-slug",
          "reasoning": "exact-level",
          "verified_by": "catalog source and probe command",
          "verified_on": "YYYY-MM-DD"
        },
        "medium": {},
        "large": {}
      }
    }
  }
}

Fill every tier object. Do not leave placeholders.

5. Preview and apply

Preview the transaction:

python3 scripts/configure.py --repo <repo-root> --plan <plan.json> --check

Show the user the selected harnesses and tier mappings. Then apply:

python3 scripts/configure.py --repo <repo-root> --plan <plan.json> --mirror-claude-skills

The mirror flag has an effect only when Claude Code is selected. It copies owned ps-* skills from .agents/skills/ into .claude/skills/ without replacing unowned directories.

6. Verify discovery

Read back .agents/pstack/models.json, the bounded AGENTS.md block, and every generated profile. For Claude Code, also verify the CLAUDE.md import and mirrored skills.

Start a new session in each configured harness. Confirm that ps-setup-pstack and ps-poteto-mode are discoverable. Run one harmless subagent task at each tier when the harness supports delegation. Report harnesses that can run P-Stack only inline.

An unchanged rerun must produce byte-identical generated files.