Setup

---
name: setup
description: "Finish installing boardless-pcb on Codex or Cursor — installs the amnesia-verifier subagent and the always-on workflow rules, which the skills installer cannot carry. Run once per project, right after installing the skills. Not needed on Claude Code, where the plugin ships both."
disable-model-invocation: true
---

# Setup — finish the install (Codex / Cursor)

`npx skills` installs skills and nothing else. Two parts of boardless-pcb are not skills:

- the **amnesia-verifier subagent** that Step 6 delegates its independent audit to
- the **always-on workflow rules** that belong in the project's `AGENTS.md`

This skill installs both from the files bundled beside it, into the project the user is working in. Run it once per project. It is **idempotent**: re-running reports what is already in place instead of duplicating it.

Respond in the user's language.

## 1. Host platform

- **Claude Code** → nothing to do. The plugin already ships the subagent *and* the anti-patching hook. Say so, point the user at `/boardless-pcb:step1-onboard`, and stop here.
- **Codex** → §2, using `assets/codex/amnesia-verifier.toml` → `<project>/.codex/agents/amnesia-verifier.toml`
- **Cursor** → §2, using `assets/cursor/amnesia-verifier.md` → `<project>/.cursor/agents/amnesia-verifier.md`
- **Anything else** → say plainly that this workflow has only been built for these three, and stop. Do not improvise a target directory.

Confirm the project root with the user before writing anything — this must land in the PCB project, not in whatever directory the session happens to start in.

## 2. Install the subagent

Copy the file for the detected platform, byte for byte. Do not summarize, reformat, or "improve" it: its wording is the whole mechanism — an independent reviewer that reads only live EDA data is what makes Step 6 catch what the main agent cannot.

- Target missing → create the directory and copy.
- Target already identical → report "already installed", change nothing.
- Target exists and differs → show the user what differs and ask before overwriting. A user may have hardened it deliberately (for example by narrowing its tool access); do not silently undo that.

## 3. Install the always-on rules

Append `assets/AGENTS-snippet.md` to `<project>/AGENTS.md`, creating the file if it does not exist. If a `# boardless-pcb workflow rules` heading is already present, leave it alone and report it — never append a second copy.

## 4. Codex only: project trust

Codex loads project-level `.codex/agents/` **only in trusted projects**. You cannot grant that for the user. Tell them: trust the project when Codex asks on first open, or add it under `[projects]` in `~/.codex/config.toml`.

Say why it matters rather than just naming the step: without trust the subagent never registers, Step 6 silently falls back to the main agent reviewing its own design, and the independent audit — the reason this workflow finds what it finds — is gone with no error message.

## 5. Report and hand off

State plainly, in one short list: what was installed, what was already present, what the user still has to do themselves (Codex trust), and one thing that is *not* installable here — **there is no anti-patching hook on Codex or Cursor**. On Claude Code a PreToolUse hook mechanically reviews file edits for downstream patching; on these platforms the same rule binds by instruction only, through the AGENTS.md snippet you just installed.

Verification the user can run themselves:

- Codex: `/skills` lists the nine step skills; the subagent appears once the project is trusted
- Cursor: the step skills appear under `/`; confirm they are **not** in the model's implicit skill list (they carry `disable-model-invocation`, but confirm it once — this flag is the only thing keeping the steps user-invoked on Cursor)

Then point them at the entry point: start a fresh session and run **step1-onboard**.