Setup

Connect the judges

!"$CLAUDE_PLUGIN_ROOT/scripts/probe.sh" 2>/dev/null || "$HOME/.claude/skills/multi/scripts/probe.sh" 2>/dev/null || ./.claude/skills/multi/scripts/probe.sh

$SCRIPTS is whatever the probe printed as scripts-dir:. The probe already ran above — read it, do not run it again. It detected everything detectable: which CLIs are installed, who is logged in, which keys exist, whether the OpenCode paid channel is available, and what other AI CLIs live on this machine. Never ask the user about anything the probe already answered.

If the line above reads Shell substitution failed instead of probe output, the session is in a git worktree whose shell gate refused the header; the plugin is fine. Run "$HOME/.claude/skills/multi/scripts/probe.sh" (or the same under $CLAUDE_PLUGIN_ROOT) yourself, as one plain command, and read scripts-dir: from that.

The user here is not assumed technical. No jargon, no walls of text, one step at a time, and never make them feel behind for not having done this already.

First contact: what this is, in three sentences

Only when little or nothing is connected yet (Claude alone, or one backend). Returning users with a working lineup skip straight to fixing gaps.

This plugin asks several different AI models to look at the same thing instead of just one — when they agree it's a real signal, and when they disagree, that disagreement is exactly what one model alone can never show you. You'll use it mostly as /multi:code-review (several models review your changes before a PR), /multi:ask (one question, several independent answers), and /multi:check-if-done (is this actually finished?). Claude is already connected; each backend below adds an independent reviewer, most at no extra cost.

Show them the one file, before anything else

Right after first contact (and for a returning user, whenever they ask to change models, endpoints or who reviews), put the config in front of them. It is theirs to edit, and the whole point is that they can. In words, then the file:

Everything multi does is set in one file, ~/.claude/multi/config.toml: which reviewers exist, which models each one tries in order, and which set runs by default. Keys are the one thing kept out of it, in providers.env, so you can share the config without sharing a key.

If the probe printed config: built-in default, there is no file yet. Run $SCRIPTS/setup.sh init — it writes the default with a comment on every field — and show the user the file it wrote (cat it, trimmed to the backends and profiles, without the long comment header). Then say in one sentence what the default does: "right now a review runs Codex, OpenCode and OpenRouter — the last one only once it has a key." If a file already exists, show that one instead.

They keep the default or change it; either way they now know where the knob is. Every later step in this skill that "writes a model" or "adds an endpoint" is an edit to this file, and you say so each time.

One question, only about what the probe can't see

The probe knows what's installed. It cannot know what accounts the user has or what they're willing to pay. Ask one question covering only the missing backends, shaped like this (adjust to the actual gaps):

To pick what to connect, tell me:

• Do you have a ChatGPT subscription? (unlocks GPT as a reviewer, no extra cost)
• Are you okay creating one free account + API key? (unlocks many model families at once via OpenRouter or a compatible service)
• Or keep it strictly free-and-local for now? (OpenCode's free models)

Then recommend an order — the fewest steps to reach at least one non-Claude reviewer (without one, /multi:code-review refuses to run; with one it works, and each further backend makes it stronger). Typical value order: Codex if they have the subscription, otherwise an OpenRouter-style key, OpenCode as the free floor, Gemini as a free extra.

Connect each chosen backend

Work through the user's picks one at a time — give one step, wait for them to do it, then the next. Never paste four install blocks in one message.

For each backend, read its reference file first and follow it:

• Codex (GPT) — references/codex.md
• OpenCode — references/opencode.md
• OpenRouter / 9router / z.ai / any compatible endpoint — references/openrouter.md
• Gemini — references/gemini.md

Everything of this plugin lives under ~/.claude/multi/: config.toml for backends, models, endpoints, profiles and timeouts (read references/config.md before editing it), providers.env for keys only. If the probe printed models-config: LEGACY, the old opencode list file is no longer read — carry its models into [backends.opencode] in config.toml and delete the file. If the probe printed config: BROKEN, fix that first: nothing runs on a broken config, and it names the key and the reason.

If the probe listed other-ai-clis, you may mention them in one sentence as detected-but-not-yet-supported. Do not improvise support for them.

Picking models

When a backend is connected but the model choice is the question (free vs paid, which pin, which fallbacks) — or the user asks "which models should I use" — offer the fork:

1. Sane default — keep what the probe picked; say in one line what that is and what it costs (usually: free, weaker, can flake).
2. "Research it for me" — read references/model-research.md and follow it: live catalogue + web search (your own model knowledge is stale), propose a pick, write it into config.toml only after the user approves.

Keys never touch this chat

Not a preference — a key that reaches this conversation is a leaked key: it is in the transcript, and if it also reached a command line it is in shell history and was visible in ps. So setup.sh set <NAME> is given without a value and run by the user, in their own terminal: it prompts, reads the key without echoing, and writes it out of sight. If a key lands in the chat anyway, say plainly it should be rotated at the provider — deleting a message does not un-leak it.

Keys live in ~/.claude/multi/providers.env, permission 600 (owner-only). On Windows/MSYS and some mounts chmod silently does nothing; setup.sh checks afterwards and warns when that happened — if it warned, repeat the warning to the user instead of claiming the file is protected.

Finish: show where things stand

Run:

$SCRIPTS/setup.sh status

This one actually verifies keys over the network (the probe above does not). Report one short block — who's connected, and what each missing one would add:

Connected: Claude, Codex, OpenRouter (endpoint: 9router)
Not yet:   OpenCode (free extra reviewer), Gemini (free extra reviewer)

/multi:code-review works — you have non-Claude reviewers. Each missing
backend is one more independent opinion, most at no extra cost.

Be honest about degradation, not alarming: missing backends are a normal starting state, not a broken install. Then stop — this skill's only job is getting the judges connected, not running a review.